Awesome Jev

Discover Jev-powered apps, developer resources, and runnable decision examples.

Jev is TypeSafe AI's System One model: give it state and typed questions, then use the returned choices, scores, and probabilities in your code. This independent community directory collects applications that use Jev, resources for developers, and small, inspectable workflows to learn from.

Explore Jev-powered apps · Share your app

Try the beta web UI at JevList. Explore the projects in this directory through a searchable web interface. We're continually improving the experience—take a look and let us know what you think!

Contents

• Explore Jev with our skill
• Start here
  • Explore practical use cases
• Official SDKs and tools
• Community projects
  • Community knowledge base
  • Apps powered by Jev
  • Browse the project directory
  • Tools and integrations
• Computer and browser use
• Starter projects
• Reference project
• Patterns and cookbooks
  • Routing and classification
  • Retrieval and verification
  • Extraction and structured data
• Model behavior and evaluation
• Contributing

Explore Jev with our skill

Use Awesome Jev Guide to understand Jev and explore solutions that fit your needs. Describe your workflow in plain language, and your coding agent will help you choose an example, adapt a project, or build a suitable starter—with guided setup along the way.

1. Install the skill. With Node.js 22+ installed, run this in your project directory:

npx skills add AppitStudio/awesome-jev --skill awesome-jev-guide

Select your coding agent when prompted. For Codex, you can add --agent codex. No TypeSafe account or API key is needed to explore ideas or run the offline examples.

2. Open your coding agent in that project and try a prompt.

New to Jev? Start here:

Use awesome-jev-guide. Explain Jev simply, help me understand what I can
build with it, and walk me through the easiest example. I don't have
an API key yet.

Have a workflow in mind? Fill in the brackets:

Use awesome-jev-guide. Today I [describe the work I do manually].
I want to [describe the result]. I use [language, framework, or tools].
Help me find the best starting point in this repo and explain what
needs adapting. Ask me anything essential that's missing.

Ready to build?

Use awesome-jev-guide to build a small starter for [my workflow] using
[my stack]. Start with a working offline demo, explain what I should
customize, then guide me through API-key setup and a first live check
when I'm ready.

3. Follow the next step together. The guide uses current TypeSafe documentation and explains what already works, what needs building, and how to handle uncertain results. When you want live access, it walks you through private key setup—never paste a key into chat.

See the full onboarding guide for installation options, demos, and setup details.

Start here

Explore what you can build — choose a real implementation by the result you want, then follow its guide, worked example, and setup path.

• Introduction - Understand Jev's state-and-questions interface and its three decision primitives.
• Official quick start - Make a first request using Python or HTTP.
• Run an example locally - Try a complete workflow with synthetic mock responses, without an account, an API key, or package installation.
• Choose a decision pattern - Match a task to Choice, Score, or Noul and define what happens when the result is uncertain.
• HTTP API reference - Check the wire format, authentication, answer fields, and error responses.

Official SDKs and tools

These resources are maintained by TypeSafe.

• Agent skill - Official docs for giving a coding agent API context and guidance when designing narrow System One / Jev decision questions (Claude Code plugin and skills.sh install paths).
• JavaScript SDK - JavaScript and TypeScript client that infers answer types from the supplied questions.
• Python SDK - Synchronous and asynchronous clients with typed answers and configurable retries.
• System One Adapter - Run a similar typed-question interface against other LLM providers for comparisons; those responses do not come from Jev.
• TypeSafe Agent Skills - Official Claude Code / skills.sh agent skill for designing typed System One / Jev decisions; docs.

Community projects

The community knowledge base turns credited articles into original implementation guides with source links, project connections and stated validation limits.

Browse the project directory for separate app and tool directories, each with full project pages covering use cases, setup, examples, limitations, and review evidence. The links below still take you directly to the upstream projects.

Checks are tied to reviewed versions; see each page and the validation scope. Mocked tests and reported live smoke checks do not establish model quality on your workload.

Apps powered by Jev

• 1 Million Emojis - Open source · Free · BYOK. Shared million-emoji canvas where TypeSafe Jev paints alongside humans from stroke context. Try app · Project guide.
• AI Model Picker - Open source · Free. macOS task-to-model picker that opens the task in your existing AI app; optional TypeSafe Jev classification. Project guide.
• AI Plays Pokémon (kotlinds) - Source available · Free source build · BYOK. Kotlin desktop app that puts an AI in front of Pokémon HeartGold on a Nintendo DS emulator: the game RAM is described as JSON, and a Jev backend picks each action from a closed list with per-option probabilities, optionally with an LLM planner in hybrid mode. Project guide.
• AI, SI, or Whatever - Open source · Free source build · BYOK. Chrome extension where TypeSafe Jev can choose AI/SI/alternating/custom webpage terminology per page. Project guide.
• AI;DR - Source available · Free. Ranked daily AI-news site and digest whose importance scoring can use Jev through AnyRouter, with a chat-model rubric as backup. Project guide.
• Albatross - Open source · Free source build · BYOK. Terminal-first coding harness (bring your own model, key or MCP server) with an experimental Jev Mode that puts TypeSafe's decision model before the main LLM. Project guide.
• AlgorithVideo (open-source desktop video clipper with optional Jev clip review) - Open source · Free · BYOK. Open-source Electron desktop app that turns podcasts, interviews and streams into captioned short clips rendered on your own computer; an opt-in Jev review step, through OpenRouter's Decisions API, scores candidate moments from transcript excerpts. Project guide.
• Anki Review-Finder - Open source · Free source build · BYOK. Alpha CLI that audits Anki flashcards through AnkiConnect with TypeSafe Jev, flagging structural defects such as answer leaks, missing context, scope problems and binary questions, then writes priority and defect tags back to Anki (never card text) and produces an HTML report. Project guide.
• Anti Slop for X (Chrome extension that folds AI-bait and bot replies on X, scored by Jev) - Open source · Free source build · BYOK. Chrome extension for x.com that pre-screens replies with local keywords, then optionally sends a post and its parent to a decision model through your own proxy worker. Jev answers whether it reads like AI engagement bait, which kind of bait it is and how strongly; replies over the threshold are folded with the score and reason shown, and mute or block only happens when you click. Inspired by slop-filter (MIT) but written independently. README in Chinese. Project guide.
• Aperture (ranked reading feed) - Open source · Free. Single-user, self-hosted intelligence feed that imports Readwise Reader documents and optional public sources (Reddit, GitHub Trending, YouTube, reports, CVEs), has Jev rate each item for quality, personal utility and actionability, and shows a ranked, explainable reading queue with summaries. Project guide.
• Aplausômetro - Open source · Free source build · BYOK. Paste a social post and watch a simulated audience react in real time with TypeSafe Jev calibrated probabilities. Try app · Project guide.
• Apparite (jev2ui) - Open source · Free source build · BYOK. Local design-mock lab: TypeSafe Jev chooses IA/anatomy; Gemini writes copy; code assembles A2UI-inspired mocks. Project guide.
• arJev - Open source · Free source build · BYOK. Daily arXiv digest engine: lexical front-line plus TypeSafe Jev confidence-gated rerank against an Obsidian vault. Project guide.
• AskJev (openconstruct) - Open source · Free · BYOK. Static single-page yes/no UI: OpenRouter Decisions + TypeSafe Jev Noul; key stays in browser localStorage. Project guide.
• ASO Workbench (App Store Optimization on your Mac, with Jev keyword clues) - Source available · Free. Mac app for App Store Optimization that tracks keyword ranks per country, estimates popularity and difficulty with labelled sources, writes and publishes store listings in every language, and measures what each change did; Jev answers narrow keyword questions that feed its popularity estimate. Project guide.
• Atmosphere (animate a rigged 3D character from a plain-language action, with Jev judging each pose) - Open source · Free source build. Web app built for the Devpost Build With AI: Basics hackathon: upload a rigged 3D character, describe an action, and an OpenAI model plans the timing, beats and poses while Jev judges and polishes each pose before the result plays back. Project guide.
• Attention Vaccine (Chrome extension that flags manipulation techniques with Jev) - Open source · Free. Chrome extension that tags posts and articles using manipulation techniques such as rage bait, easy-money pitches, impersonation notices, dehumanizing language and unsupported assertions, without judging viewpoints or fact-checking. Project guide.
• Audience Moderator (Android counter for people and raised hands, with an optional Jev round insight) - Open source · Free source build. Android helper for event moderators: counts people and raised hands on-device (MediaPipe or YOLOv8n-pose), runs timed question rounds, picks a random audience member, and can optionally send round results, or one photo you confirm, to Jev for a short insight and a second count; camera pictures are never saved and there is no face recognition. Project guide.
• Auto-Oui - Open source · Free · BYOK. Chrome extension (French UI) that watches a tab you enable and answers "Oui" — or clicks the approve button — when an AI agent on supported sites stops to ask for permission; detection is local rules by default with no network calls, and an optional Jev mode with your TypeSafe key judges ambiguous wording against a probability threshold (0.90 default). The README warns to keep it off for payments, messages, deletions and production actions. Project guide.
• axo-launcher - Open source · Free source build · BYOK. Minimalist Android launcher where TypeSafe Jev settles ambiguous quick-capture phrases (Portuguese docs). Project guide.
• Baton (notch app) - Open source · Free source build · BYOK. macOS notch app that alerts you when AI chats finish or need you; TypeSafe Jev decides which. Project guide.
• Beam (Omarchy launcher with optional Jev intent search) - Open source · Free. One-box launcher plugin for Omarchy that launches apps, runs Omarchy menu actions, calculates and converts inline, and opens web searches and URLs; with an optional Jev key it also finds actions by meaning, so typing screen warmer finds Nightlight. Project guide.
• Beat Jev - Open source · Free source build · BYOK. Penalty shootout against TypeSafe Jev on Render Workflows with Python and TypeScript examples. Project guide.
• beebots - Open source · Free source build · BYOK. Three OKX paper-trading bees; every trade decision via TypeSafe Jev (OpenAI only for bee design/portraits). Not financial advice. Project guide.
• Better Email Filter (local Gmail triage with Jev that learns from your corrections) - Source available · Free source build · BYOK. Local command-line Gmail triage: Jev sorts messages into Spam, Ignore or Important, sends uncertain ones to Review, and the labels you correct in Gmail become examples for future classifications. Project guide.
• betterx - Source available · Free source build · BYOK. Chrome/Brave extension that classifies posts and replies on x.com with TypeSafe Jev as you browse — seven questions per post in one ~150–250 ms request (tone, post type, argument quality, racist, contempt for a nationality or immigrants, antisemitic, sexually explicit) — and labels, dims, blurs or hides them by your rules. Project guide.
• Bilibili noise filter for Firefox (JEV) - Open source · Free source build · BYOK. Firefox (and Chrome) port of littlewindy123/jev-bili-filter: toggles for spoilers, trolling, ads and partisan chants, plus your own one-sentence rules, filter Bilibili comments and danmaku with TypeSafe Jev; adds local keywords that work without a key, per-item Canvas danmaku filtering, a block log and usage tracking. Project guide.
• Booksmith (Calibre ebooks to Plex audiobooks, with Jev-planned pauses) - Source available · Free source build · BYOK. Early Windows-first app that turns DRM-free Calibre ebooks into chaptered audiobooks narrated locally with Kokoro and imports them into Plex; optional Jev annotation proposes sentence pauses, reviews them in paragraph context and suggests pronunciations for ambiguous words. Project guide.
• BOOLS (Breakout played live by Jev in the browser) - Open source · Free · BYOK. Browser Breakout game where a System One decision model steers the paddle: while the ball falls the page sends its state and one Choice question over eight zones, and the paddle moves to the pick; on-screen stats show latency, accuracy by distance and per-zone probabilities. Project guide.
• Bot Derby (Jev and OpenAI Decisions compete in a paper-trading experiment) - Open source · Free. Public paper-trading experiment: Ferdi (TypeSafe Jev) and Bodi (OpenAI Decisions) each get $50,000 of simulated cash, the same stock evidence and the same legal action menu for five large US stocks, and a public scoreboard tracks how their portfolios diverge. Project guide.
• buildornot (BUILD, MAYBE or DON'T BUILD verdicts on side-project ideas, judged by Jev) - Open source · Free source build · BYOK. Command-line tool that judges a side-project idea: it searches GitHub and Hacker News for prior art, has Jev rate pain, reach, feasibility and deal-breakers and how likely each match does the same job, and prints BUILD, MAYBE or DON'T BUILD with the reasons. Project guide.
• BUTLR.ai (Windows desktop assistant that operates the focused app, choosing each UI action with Jev) - Open source · Free source build · BYOK. Focus any Windows app, press Ctrl+Win and type what you want done. BUTLR.ai reads the app's accessible controls through Windows UI Automation, asks Jev (through Vercel AI Gateway) to pick the next action, then types, clicks and checks the result. Steps such as Submit, Send, Pay, Delete or Install always wait for your approval, password fields and card numbers are never captured, and password managers are blocked. Project guide.
• B站降噪 · JEV (Bilibili noise filter) - Open source · Free · BYOK. Original Chrome/Edge extension (Chinese, with English README) that filters Bilibili comments and plain-text danmaku on the page with TypeSafe Jev: four toggles — spoilers, trolling, ads and fandom flame wars — plus one-sentence custom rules, no blocklists or prompts to maintain. Upstream of the listed Firefox port. Project guide.
• Call Coach - Open source · Free source build · BYOK. Local live sales-call coach: TypeSafe Jev picks next actions and buying stage from transcript turns (mic or sample call). Project guide.
• Call Coach by Darlison (live interview checklist on macOS, ticked by Jev) - Open source · Free source build. Live checklist for interviews and discovery calls: whisper.cpp transcribes both sides on your Mac, Jev ticks off each question once it is answered, tracks which listed question you just asked and flags answers worth a follow-up, and an MCP server lets your AI add follow-ups. Project guide.
• Capture (sgaabdu4) - Open source · Free source build · BYOK. Private Mac voice diary: local Parakeet transcription, TypeSafe Jev sorting, Notion after approval. Project guide.
• CareerSurge (free desktop app that finds job postings and applies for you locally, with Jev only picking among values it already found) - Open source · Free source build · BYOK. Local desktop app that searches Greenhouse, Lever and Ashby listings through your own Apify account, shows a batch to approve, and fills and submits each form in a headless Chromium. It does no captcha solving or made-up answers; Jev, if you add a key, only picks among values the app already found (resume parsing, label mapping, question classification), and anything unresolved waits in 'Needs you'. Project guide.
• ChatJev by JamesANZ (ask Jev to judge a document or web lookup in plain language) - Open source · Free source build. Web app where you upload a document or ask about the web: a fast LLM turns your question into Jev yes/no, choice or score judgments, can attach web excerpts as evidence, and then restates only Jev's numbers, with a verdict card showing each judgment. Project guide.
• ChatJevPT - Open source · Free source build · BYOK. Chat app that makes TypeSafe Jev 'generate' an answer one character at a time by asking the same choice question over and over (next character or END), shows each letter's top-five probabilities, lets Jev grade its own answer, and meters cost; several 'model levels' compare naive and fuller pipelines. Project guide.
• ChatJump - Open source · Free source build · BYOK. Menu-bar launcher for your AI chats: press Control-Option-Space, type what you remember, and Jev ranks open Claude, ChatGPT, Codex and Chrome chat tabs by meaning, then opens the right one. Project guide.
• CheckMyBrandName (free brand-name checker where rules gather evidence and Jev judges what rules can't) - Closed source · Free. Free brand-name screen for spelling, pronunciation, distinctiveness, meanings in six languages and domain availability, with Jev judging what rules can't (site claim); no sign-up, also an MCP server. Try app · Project guide.
• Cheese Toast - Open source · Free. Classic teapot-in-space arcade game in the browser; TypeSafe Jev moderates score names and can play as a bot. Project guide.
• Chess with Jev - Source available · Free source build · BYOK. Browser chess where your opponent is TypeSafe Jev: every move is a short chain of choice questions (focus, piece, destination) whose options are the legal moves, so illegal moves cannot be expressed; you can read each answer chain and rewrite the questions in an in-app workflow editor. Project guide.
• Clasificador de noticias regionales con Jev - Open source · Free source build · BYOK. Spanish-language Windows workflow that turns a monthly multi-portal news Excel into a reviewed list of Arequipa economic news, using Jev to classify headlines before scraping. Project guide.
• Clean Code Review - Open source · Free source build · BYOK. Hosted/source PR reviewer: TypeSafe Jev judges files on Clean Code questions; Luna writes evidence-first prose (MCP included). Try app · Project guide.
• Clef homelab decision engine - Open source · Free source build. Self-hosted homelab dashboard where open Clef-flash (GGUF on a 6 GB laptop GPU) answers System One questions over llama.cpp's /v1/systemone endpoint — the same request shape as Jev — to triage Gmail, filter news, watch ~2,350 company job boards for internships and push matches to your phone through ntfy. Project guide.
• Command Space (macOS Spotlight replacement with an agent, local decisions and optional Jev) - Open source · Free · BYOK. Launcher that replaces ⌘Space: apps, documents, paths, math and settings, plus an AI agent for questions and tasks and push-to-talk voice. Quick typed decisions run on a local model by default, with TypeSafe's Jev as an optional hosted backend. Project guide.
• Computer Use (lite) by ajokhai (Mac computer use for any AI, with an optional Jev mode) - Source available · Free · BYOK. Menu bar app and native MCP server that let an AI operate the whole Mac, not just a browser: Claude, ChatGPT, Gemini, OpenRouter or a local model plans the steps, and in its Jev modes TypeSafe Jev picks the next action and element from the app's accessibility tree. Project guide.
• Conclave (local multi-model council app with Jev routing and answer checks) - Open source · Free source build · BYOK. Local Windows app that sends one question to a council of up to 13 AI providers, has a chosen model synthesise their answers blind and an optional model from another company audit the result; optional Jev checks route questions, flag refusals or off-topic answers and test draft claims, in shadow (record-only) mode by default. Project guide.
• Covan (local-first personal agent desktop app with Jev reply decisions) - Open source · Free source build · BYOK. Local-first personal agent desktop app with memory, tools, workspaces and scheduled tasks, where several bots can chat with you or in groups; an optional Jev strategy decides whether each bot should reply in a group or follow up, with low-confidence answers falling back to the LLM strategy. Project guide.
• CrawlHive - Open source · Pricing unverified. Memecoin signal crawler swarm: TypeSafe Jev rates every crawler find in ~0.5 s before safety scans and optional Grok review. Project guide.
• Crush Monitor - Open source · Free source build · BYOK. Local WeChat-style chat analyzer: TypeSafe Jev labels emotion/intent, scores affinity, and rates replies (BYOK). Project guide.
• Crush Monitor with Jev - Open source · Free source build · BYOK. WeChat affinity analyzer (MIT fork of Crush Monitor) powered by TypeSafe Jev judgments. Project guide.
• CryptoCadence - Open source · Free source build · BYOK. Paper-trading/backtest crypto stack with calibrated TypeSafe Jev decisions, hard risk vetoes, and a tamper-evident journal. Project guide.
• Danmaku spoiler filter (JEV) - Open source · Free source build · BYOK. Zero-dependency Chrome extension (plus CLI) that has TypeSafe Jev judge whether Bilibili danmaku comments are spoilers and filters the hits, batching comments per request with a local cache; built to add other danmaku sites later. Docs in Chinese and English. Project guide.
• DeciFlow - Open source · Free source build · BYOK. Self-hosted log triage: groups log events into chunks and classifies each one as Normal, Warning, Error, Security or Other with urgency and a needs-a-ticket probability, using self-hosted Laya by default or TypeSafe Jev; batch scans, live follow and a dashboard with probabilities. Project guide.
• Decision Lab - Open source · Free source build · BYOK. Visual System One playground to build decisions, compare Jev/OpenJEV/Laya, and explore what-if changes. Project guide.
• dicechess-bot-jev - Open source · Free source build · BYOK. Experimental Dice Chess webhook bot: TypeSafe Jev picks among legal turns without heuristic pre-ranking (AGPL-3.0). Project guide.
• Did They Answer? - Open source · Free. Web app that scores how thoroughly the UK government engages with the petitions it answers: Jev rates each response against a published rubric, with a hand-labelled gold set split into tune and holdout to keep prompt changes honest. Project guide.
• Doom or Bloom - Open source · Free · BYOK. Map your AI worldview through adaptive interview questions powered by TypeSafe Jev. Try app · Project guide.
• dotpals - Open source · Free · BYOK. Desktop pal and notch showing what coding agents did (tests read from output, faked passes caught); an opt-in check sends unclear test runs to TypeSafe Jev via @typesafe-ai/sdk or to local Laya. Project guide.
• DSB (trust-first Android AI assistant with a Jev command gate) - Open source · Free source build. Open-source Android assistant built on DeepSeek Harness that operates apps on a background virtual display of your own phone, with execution gates, a task ledger and audit logs; an optional Jev gate classifies suspicious shell commands before they run. Project guide.
• Dungeon of Jev - Open source · Free source build · BYOK. Turn-based browser dungeon crawler where Jev picks which champion to send, which form the shapeshifting boss opens with, and every enemy move, as single Choice questions. Project guide.
• DurableMMO (cozy multiplayer farming game on Cloudflare where villagers can make decisions with Jev) - Open source · Free source build · BYOK. Multiplayer farming game in the valley of Mossbrook: farm, mine, fight light monsters and chat with ten villagers. Villager decisions come from rules, optionally TypeSafe Jev, while Workers AI writes their lines. Runs on Cloudflare Workers and Durable Objects with a PixiJS client. Project guide.
• Elefante (desktop multi-agent workspace with an API hub and browser harness, where Jev can route channel messages) - Open source · Free source build · BYOK. Electron workspace (bilingual English/Portuguese README) for running many AI agents on any engine, from OpenRouter, DeepSeek, Ollama, Groq, OpenAI and Anthropic to CLI harnesses such as Claude Code and Codex, with a native API hub and browser harness. Jev is an optional router that decides which agent takes a message. Project guide.
• Elmo (open-source AI visibility tracker where Jev judges brand mentions and sentiment) - Open source · Freemium · Commercial · BYOK. Open-source answer-engine-optimization platform: runs your tracked prompts against ChatGPT, Google AI Mode, Gemini, Perplexity, Copilot and model APIs, then reports visibility, share of voice, citations and sentiment. Jev judges whether each passage really mentions your brand or a competitor and how it is framed, keeping the passage each judgment rests on. Project guide.
• Elo Jev 4 Investment Ideas - Open source · Free source build · BYOK. Local Elo tournament where TypeSafe Jev judges pairwise investment-idea matchups per criterion. Project guide.
• Emc² (desktop email client with Jev sorting) - Open source · Free. Desktop email client for Windows (macOS build from source) with a bundled mail service and encrypted SQLite storage; Jev sorts mail into Important and Other and gives each row a fixed label such as "Needs a reply". Project guide.
• enigma-jev - Open source · Free · BYOK. Software Enigma/Bombe where TypeSafe Jev ranks cribs and judges German plaintext (hosted + CLI). Try app · Project guide.
• extynt (After Effects plugin and MCP server: plain-language motion design with optional Jev checks) - Open source · Free. Native After Effects plugin for macOS and Windows: describe a motion-design change in plain language, a model proposes a typed plan, and extynt validates, previews, applies it in one undo group and verifies the result; Claude Code and Codex can drive the same project over MCP, and small closed questions can optionally be answered by Jev when it is confident. Project guide.
• Eyedot · 点睛 - Open source · Freemium · Commercial · BYOK. Turns your own study material (PDF, Word, text) into exams that grade themselves point by point: grading is decomposed into atomic checkable questions whose probabilities are combined in code, with TypeSafe Jev as the decision engine when TYPESAFE_API_KEY is set (general-LLM judge and a labelled lexical demo engine as fallbacks), plus cited follow-up answers and spaced-repetition review. Project guide.
• Facet - Open source · Free source build · BYOK. Photo culling and sorting app for photographers: an AI model checks every photo against built-in and your own labels (blurry, eyes closed, the couple kissing), faces are grouped locally, and you filter and sort into folders with undo. Jev or OpenAI vision can be chosen. Project guide.
• Fallacy Lens - Open source · Free source build · BYOK. Chrome extension: select text, right-click and Jev (via OpenRouter, your own key) scores it for 69 logical fallacies, showing the highest-scoring ones first with explanations. Project guide.
• Family Feud Live (multiplayer party game whose optional AI host uses Jev to judge answers) - Open source · Free. Jackbox-style multiplayer Family Feud fan game: the board runs on a TV, players join from phones by QR code, and an optional AI host writes surveys on any topic, speaks in character and judges fuzzy answers, with a local matcher first and Jev as the decision model for the guesses it cannot decide. Project guide.
• FilmHunch - Open source · Free source build · BYOK. Describe a movie any way you like; TypeSafe Jev ranks films against your exact sentence (self-hosted, BYOK). Project guide.
• fin (agent-first double-entry finance ledger where Jev proposes matches and categories) - Open source · Free source build · BYOK. Gives your AI agent a trustworthy place to manage your money: a double-entry ledger with an agent-first CLI and a web view a human can check at a glance. Code classifies what it can first (learned UPI IDs, transfers, card patterns, who owes whom); Jev handles the remaining closed-set judgments such as 'same payment?' and categories, and only proposes; it never writes. Project guide.
• Find the Right API (Orthogonal × Jev) - Closed source · Freemium · Commercial. Hosted goal→API discovery: product copy says Jev searches ~1,000 Orthogonal endpoints and returns priced matches with Orthogonal run snippets (launch). Try app · Project guide.
• FinDoc (loan-file copilot that checks a mortgage packet against its documents, with Jev classifying and splitting pages) - Open source · Free source build · BYOK. Web app for mortgage loan officers: upload a borrower's PDFs, get an automatic review of the application against pay slips, statements and title documents, record a decision on every flag and ask follow-ups with page citations. Jev labels each page's document type and detects document boundaries during ingestion; the app never approves or rejects loans. Project guide.
• fine-file (Windows file explorer with optional Jev risk badges) - Open source · Free · BYOK. Korean-language Windows file explorer with folder sizes, instant search from an NTFS MFT scan and live updates, plus optional AI clean-up plans where Jev adds fast safe, caution or risky badges and a second risk check on each step. Project guide.
• Finney (finance agent for hackathon directors, with Jev reviewing agent traces) - Open source · Free source build · BYOK. Event finance workspace for hackathon directors that keeps source evidence, budget commitments, decisions and reports together in Supabase with a Next.js front end; OpenCode and Hermes do scoped agent work, Jev asks four yes/no questions about each finished trace (unsupported completions, missing payer evidence, needless clarifications, stale links), and payments run only on Base Sepolia test funds with exact-action approvals. Project guide.
• Flash (YouTube Q&A) - Source available · Free · BYOK. Ask a YouTube video a question and jump to the answer; TypeSafe Jev scores transcript passages (BYOK). Project guide.
• FlowMesh - Open source · Free. Windows voice-workflow app built for the Wispr No Hands challenge: dictate into the command box, review an exact plan, then approve; Jev picks which registered routine (focus, meeting, research or none) fits, and a local runtime runs only allowlisted actions with receipts. Project guide.
• Fly x Jev - Open source · Free source build · BYOK. Browser demo of a MaleCNS fruit-fly connectome robot where TypeSafe Jev chooses which descending neuron fires next. Project guide.
• Focus - Open source · Free source build · BYOK. Browser extension that classifies domains as productive or distracting with TypeSafe Jev via OpenRouter Decisions and blocks distracting navigations. Project guide.
• focuslog (study-time tracker for Hyprland and Omarchy that can ask Jev whether a window supports your goal) - Open source · Free source build · BYOK. Local study and time-use tracker for Hyprland: it samples the focused window every few seconds, labels activity as study, waste or neutral with local rules and optionally Jev, and shows totals in a loopback dashboard and an Omarchy bar widget. Project guide.
• Fotocopiatrice - Open source · Free source build · BYOK. Italian Camera amendment explorer: code dedupes identical texts; TypeSafe Jev judges attributes and near-duplicate pairs (static site). Try app · Project guide.
• FrameFind - Open source · Free. Local Python web app that samples frames from an uploaded clip or public video URL, asks Jev image questions to find moments matching a plain-language description, and lets you play or export the matching snippets as MP4. Project guide.
• GenOffice (open-source AI office suite whose file search can be reranked by Jev) - Open source · Free · BYOK. Free, open-source desktop office suite for Word, Excel, PowerPoint, PDF, Markdown and HTML files with a built-in AI panel, plus a genoffice CLI and agent skill so Claude Code, Codex and Cursor can edit real .docx/.xlsx/.pptx files locally. Its home-screen file search can switch on Jev reranking, which reads excerpts of the top 20 local hits and moves the document that actually answers the query to the top. Project guide.
• GitMemoMark (Chrome extension that saves X posts and web pages to your GitHub repo as Markdown, with optional Jev folder sorting) - Open source · Free source build · BYOK. Chrome extension (Manifest V3) that saves X posts, web articles, selections, links and images to a GitHub repository you own as Markdown, with folders, tags and notes. Folder choice goes domain rules first, then Jev if you turned it on, then an unsorted folder. Japanese UI and docs; developer preview, not on the Chrome Web Store. Project guide.
• Gmail Classifier - Open source · Free source build · BYOK. Local Gmail triage: TypeSafe Jev classifies threads; confirm-then-apply labels with undo, plus phishing/unsubscribe helpers and Chrome sidebar. Project guide.
• Gmail notifications with Jev (Safari) - Open source · Free source build · BYOK. Free, open-source Safari extension for Gmail on Mac: write in plain language which emails deserve your attention, and Jev decides which new messages trigger a floating alert. Interface and docs are in Spanish. Project guide.
• gmessages-filter (Google Messages Web auto-filter with one-call multi-category Jev classification and KDE alerts) - Open source · Free source build · BYOK. WebExtension plus local sync daemon for Google Messages Web: it blocks spam, cleans up messages by age, classifies non-contact messages against all your AI categories in one OpenRouter call to typesafe/jev-router with disk caching and per-category age thresholds, and raises KDE Plasma 6 alerts for VIP senders. Project guide.
• Goatside - Open source · Free source build · BYOK. Self-hosted finance tracker where TypeSafe Jev files plain-words entries by category and direction. Project guide.
• Goutoujunshi Jev Chat - Open source · Free source build · BYOK. WeChat desktop companion: OCR + optional TypeSafe Jev strategy before reply drafts (manual send). Project guide.
• Hearth - Open source · Free source build · BYOK. Local Chrome agent that searches four rental marketplaces with TypeSafe Jev action choice and returns a shortlist (read-only). Project guide.
• Heihua Translator - Open source · Free source build · BYOK. Workplace jargon (职场黑话) reader: guess first, then reveal TypeSafe Jev’s typed reading with confidence bands—including “I don’t understand.” Project guide.
• Heliotrope (carbon-aware load scheduler with Jev load intake) - Open source · Free source build. Scheduler that shifts EV charging, water heating, laundry and pumps into low-carbon hours without missing deadlines, using OR-Tools; Jev turns plain-words load descriptions into typed load specs and ranks how essential each appliance is. Project guide.
• Hermes Focus (Mac app that keeps you on task in 25-minute blocks, judged by Jev on your own keys) - Open source · Free source build · BYOK. Mac focus app: type what you are about to do, start a 25-minute block, and the app reads the front window once a second and asks Jev whether you are on task; a nudge appears only after about twenty seconds of confident drift, with one "I'm on task" button to correct it. Each block ends with a score out of 1,000 weighed against a rubric for your goal. No server or account. Project guide.
• HookMeter - Open source · Free source build · BYOK. Chrome extension that scores social drafts as you type with TypeSafe Jev (curiosity/arousal/pattern/clickbait) via optional worker/backend. Project guide.
• Hx - Open source · Free source build · BYOK. Clinical note checklist that ticks items against clauses from the note using TypeSafe Jev (never generates text). Project guide.
• Hyperclip (macOS video clipper with Jev clip scoring) - Open source · Free source build. Native SwiftUI app that turns long videos into short captioned clips on your Mac: it downloads with yt-dlp, transcribes on device, finds clips with built-in scoring, a Jev-compatible decision model or an LLM, frames the speaker and burns in word-by-word captions. Project guide.
• Hémicycle - Open source · Free source build · BYOK. French civic vote quiz whose pipeline uses TypeSafe Jev to classify votes and check summary neutrality (French docs). Project guide.
• iMessage Bridge (Mac menu bar app that gives AI agents approved access to your iMessages, with Jev screening incoming texts) - Open source · Free · BYOK. Menu bar app that lets agents such as Claude Code or Grok Bot read and, if you allow it, send iMessages through the app instead of touching the Messages database. Each agent is approved with a pairing code and gets its own rules, every send can require confirmation, and incoming texts that try to instruct your agent can be screened by a local model and, optionally, Jev. Project guide.
• Inbox Triage - Open source · Free source build · BYOK. Local Gmail labeler: TypeSafe Jev (default) or other providers answer focused questions; code only adds triage labels. Project guide.
• Inbox Warden - Open source · Free source build · BYOK. Self-hosted agent that labels and cleans Gmail / Google Workspace inboxes with company-specific labels, using Jev as the first pass into existing labels and Claude for the rest. Project guide.
• inbox-jev - Open source · Free source build · BYOK. Self-hosted Gmail triage: every few minutes it asks Jev which category each new email belongs to, then labels, archives, unsubscribes from verified marketing and cold outreach, moves scams to Spam and leaves uncertain mail in the inbox for review. Project guide.
• INployed - Open source · Free source build · BYOK. Windows desktop app with a cloud pipeline that collects job postings, ranks them against your background, writes grounded one-page résumés, and fills and sends queued applications; Jev makes the yes-or-no calls in scoring and reads each application page when it is on. Project guide.
• IntentSQL - Open source · Free source build · BYOK. Local NL→SQLite playground with inspectable Jev/System One plans and guarded writes. Project guide.
• Jagent (jev-agent.com) - Closed source · Freemium · Commercial. Independent hosted Jev playground and gateway with no-key trial runs, games where Jev picks every move, and a Jev API on its own key (System One, rulesets, Decisions format, MCP); operator-submitted, not affiliated with TypeSafe. Try app · Project guide.
• Jarvis (Discord voice assistant for game nights, with Jev deciding whether a follow-up was meant for the bot) - Open source · Free source build · BYOK. Say "Hey Jarvis" in a Discord voice channel, ask out loud, and hear the answer in the call. Each speaker gets their own wake-word detector, answers start speaking before they are finished, and per-game patch notes keep answers current. After a reply the person can carry on without the wake word; with a TypeSafe key, Jev checks whether each follow-up was meant for the bot or a friend, so it stays quiet when unsure. Answers come from Qwen on Cerebras. Project guide.
• Jaste - Source unverified · Pricing unverified · BYOK. Mac clipboard beta with a Direct Jev mode for selecting a saved text value that fits the focused field; BYOK applies to that optional mode. Project guide.
• Jeff - Closed source · Free. Bookshelf you can ask: TypeSafe Jev ranks shelf items with per-item Noul search and Choice sorting; visitors upload Goodreads CSVs (operator pays capped Jev spend). Project guide.
• jeff (Discord bot) - Open source · Free source build · BYOK. Discord bot built with botbox where every feature is a typed Jev question: settle debates, rank lists, tag forum posts with the forum's own tags, answer FAQ repeats, and track unanswered help questions. Project guide.
• JEV (marcosdeaza) - Open source · Free source build · BYOK. macOS voice-control exploration: TypeSafe Jev decides intent and target while an Anthropic planner runs continuous sessions. Project guide.
• Jev 2048 - Open source · Free · BYOK. Instrumented 2048 web lab: every move is a TypeSafe Jev Choice with live probabilities (hosted trial or BYOK). Try app · Project guide.
• Jev abstract art playground (turns a prompt into abstract art through 80+ parallel Jev decisions) - Source available · Free. Web playground that turns a natural-language prompt into abstract art by asking Jev typed questions about 80+ renderer characteristics (palette, composition, form) in parallel, rechecking uncertain choices, then rendering with canvas-sketch. It compares latency and cost against a reasoning model and against generating the art directly with AI. Project guide.
• Jev Amazing (macOS WeChat sidebar for witty reply ideas, ranked by Jev) - Open source · Free source build. Native macOS side window for WeChat: one click reads the visible chat (text, images, stickers), a generative model drafts replies in five tones, and Jev ranks them; you edit and send yourself. A derivative of the Jev Chat assistant for macOS. Project guide.
• Jev Anti-Spam Bot - Open source · Free source build · BYOK. Self-hosted Telegram bot that deletes high-confidence spam using TypeSafe Jev Noul signals, with fail-open errors. Project guide.
• Jev Arena - Open source · Free source build · BYOK. Live head-to-head demos of TypeSafe Jev vs Claude Opus on trading, Tetris, lead racing, and review triage via OpenRouter. Project guide.
• Jev as LLM - Open source · Free · BYOK. Experimental web chat: TypeSafe Jev picks the next word via Choice questions (hosted replay demos or BYOK OpenRouter). Try app · Project guide.
• Jev Asks Until Sure - Open source · Free · BYOK. Twenty-questions web game: TypeSafe Jev keeps asking until calibrated confidence crosses a threshold (hosted demo or BYOK). Try app · Project guide.
• Jev Assistant by jbellsolutions (Mac voice assistant that works in Chrome, on your Mac and in cloud browsers) - Open source · Free. Voice assistant preview for Apple Silicon Macs: LiveKit with GPT-Live 1 handles conversation, GPT-6 Luna does reasoning and task execution, and Jev makes the fast judgments (routing, next actions and checks) for tasks in your Chrome, on your Mac, in a Browser Use cloud browser or on an optional Orgo computer, with approvals before sends and changes. Project guide.
• Jev Call Screener - Open source · Free source build · BYOK. Self-hosted call screening: TypeSafe Jev classifies caller transcripts; Go policy forwards or rejects (Twilio adapter; fail-open defaults). Project guide.
• Jev Cards - Open source · Free. Visual quiz game where you race Jev, TypeSafe's vision decision model, on the same picture cards: tap the right thing, pick the answer, with Sprint, Versus, Streak and live multiplayer modes and server-checked leaderboards. Project guide.
• Jev Chat Assistant - Open source · Free source build · BYOK. Android overlay: Accessibility reads chat; TypeSafe Jev (OpenRouter) judges intent/danger and ranks fill-only reply candidates (WeChat first). Project guide.
• Jev Chat Windows - Open source · Free source build · BYOK. Windows WeChat side panel: local OCR + TypeSafe Jev (OpenRouter) ranks fill-only reply candidates; send stays manual. Project guide.
• Jev Check (browser extension) - Open source · Free · BYOK. Chrome and Firefox side-panel extension that asks the page you are reading your own question sets (yes/no, choice, score, multi-pick) and shows probability-ranked answers from Jev, optionally judged against your uploaded resume or checklist. Project guide.
• Jev Chess - Open source · Free · BYOK. Watch TypeSafe Jev play chess vs OpenRouter LLMs, Stockfish, or you—with live probabilities and saved games (hosted demo or local). Try app · Project guide.
• JEV City - Open source · Free source build · BYOK. Illustrated isometric city simulation where every citizen decision (routine, shelter, help, restore power…) comes from TypeSafe Jev, with the full choice distribution shown on click. Project guide.
• Jev Civilization - Open source · Free source build · BYOK. Strategy game where TypeSafe Jev chooses tribe actions while you reshape the environment. Try app · Project guide.
• Jev Cloud Quiz (Japanese web demo where Jev guesses whether a feature name is AWS, Azure or Google Cloud) - Open source · Free. Japanese demo: pick one of 30 cloud feature names with brand words removed, and Jev returns a probability for AWS, Azure and Google Cloud, drawn as three bars. Ambiguous names show visibly split confidence. Project guide.
• Jev Column Race - Open source · Free source build · BYOK. Labels 1,000 app reviews in parallel: TypeSafe Jev typed questions vs Gemini JSON, with free replay (hosted or local). Try app · Project guide.
• Jev Companion (read-only WeChat conversation companion with typed Jev judgments) - Open source · Free source build · BYOK. Native macOS and Windows companion that reads the text of the current WeChat window (or pasted text) and places Jev judgment cards under each message: communication needs, next actions and conversation tension, plus optional reply drafts from a local OpenAI-compatible model; it never sends messages. Project guide.
• Jev Content Filter (hide classified ads that don't match your plain-language criteria) - Open source · Free source build. Browser extension for classified-ad and auction sites: you write what you want in plain Czech or English (price range, model year, condition), it asks Jev one yes/no question per ad in batches of ten, and collapses or dims the ads that don't match; works on Bazos.cz/.sk/.pl/.at today, with a modular design for other sites. Project guide.
• Jev Content Guard - Open source · Free source build · BYOK. Chrome extension that scans web pages for fraud, advertising, AI slop, spam, clickbait, infobusiness, and toxicity using TypeSafe Jev with configurable two-tier thresholds. Try app · Project guide.
• Jev Decision Lab - Open source · Free. Browser workbench for designing typed Jev questions, then trying them in the TypeSafe Playground. Try app · Project guide.
• Jev demos - Open source · Free source build · BYOK. Six local side-by-side TypeSafe Jev demos with Claude/Kimi switcher and simulated mode without keys. Project guide.
• Jev demos (mani-aiml) - Open source · Free source build · BYOK. Self-contained Jev/System One demo folders behind Agentic Enterprise videos (distinct from mayank953/Jev). Project guide.
• JEV Document Classification - Open source · Free source build · BYOK. Local-first folder filer: TypeSafe Jev (Vercel AI Gateway) chooses category/confidentiality/injection/subject; audit preview and undo. Project guide.
• Jev Fighter - Open source · Free · BYOK. Stick-figure fighting game and benchmark where Jev and other models fight, with replays, decision logs and Elo. Project guide.
• Jev Fit Scorer - Open source · Free source build · BYOK. Chromium extension (plus a dependency-free Node CLI) that scores the job posting you have open against your saved resume with one TypeSafe Jev request — fit 0–4 shown out of 10 with confidence, the single strongest factor, and optionally how likely the employer sponsors US visas — with cached results and per-call time and cost. Project guide.
• Jev Focus (MateusRogien) - Open source · Free source build · BYOK. Chrome extension: keep YouTube for focus music/learning while TypeSafe Jev hides the rest (distinct from Focus / jev-focus-guard). Project guide.
• Jev Focus Guard - Open source · Free source build · BYOK. Chrome MV3 extension: TypeSafe Jev classifies DOM candidates as ads/distractions and hides them (metadata-only state). Project guide.
• Jev for Chrome - Open source · Free source build · BYOK. Chrome extension that drives your current tab with TypeSafe Jev (community port of jev-ultrafast); separate text model for typing. Project guide.
• Jev for Gmail - Open source · Free source build · BYOK. Chrome extension: TypeSafe Jev priority badges on Gmail Primary with local redaction and patient-mail locks (no Gmail writes). Project guide.
• Jev Form Fill - Open source · Free source build · BYOK. Chrome extension: TypeSafe Jev proposes form field fills from clipboard text with review, readback, and undo (MIT). Project guide.
• Jev Fraud Shield - Open source · Free source build · BYOK. Explainable card-fraud triage demo/library: per-factor TypeSafe Jev judgments with a self-hosted dashboard (honest about prior failed iterations). Project guide.
• Jev FSD - Open source · Free source build · BYOK. OSM driving sim where TypeSafe Jev picks maneuvers; decision inspector + benchmark (demo only, not real AV). Project guide.
• Jev Grand Prix - Open source · Free source build · BYOK. Local F1 race: TypeSafe Jev picks line and pedals; code steers and plans next laps. Project guide.
• Jev Guess Who - Open source · Free source build · BYOK. Server-authoritative Guess Who vs TypeSafe Jev with 24 SVG portraits and analytics. Project guide.
• Jev Inbox - Open source · Free source build · BYOK. Chrome extension: TypeSafe Jev (or Vercel AI Gateway) ranks unread Gmail by criticality/urgency and your labels (list metadata only). Project guide.
• Jev Inbox Queue - Open source · Free source build · BYOK. Local inbox action queue: seven parallel TypeSafe Jev questions per thread; Python policy owns To do / Check / Filtered. Project guide.
• Jev Investment Forecast - Open source · Free source build · BYOK. Chrome extension: TypeSafe Jev forecasts investment potential across 27 page categories plus amount/duration choices. Try app · Project guide.
• Jev Kitchen - Closed source · Free. Web demo: TypeSafe Jev judges which sticker ingredients belong to a named dish or cocktail (issue #399). Try app · Project guide.
• Jev Lab (BrendanH18) - Open source · Free source build · BYOK. Local interactive workbench with demos for TypeSafe Jev System One (distinct from jammaru/jev-lab). Project guide.
• Jev Linux (voice desktop controller) - Open source · Free source build · BYOK. Voice controller for Ubuntu: local Moonshine speech recognition, a Gemini planner that splits your command into steps, and Jev choosing the desktop action for each step (open apps and sites, click, type, scroll, create notes). Project guide.
• Jev Mail (Gmail extension) - Open source · Free source build · BYOK. Open-source Chrome/Brave extension that classifies Gmail with TypeSafe Jev — category, priority, important, needs-reply and junk as separate narrow questions — then applies Jev/... labels and previews safe, Trash-only cleanup with Undo; runs in the browser with your TypeSafe or OpenRouter key. Project guide.
• Jev Mail Classifier - Open source · Free source build · BYOK. Classifies IMAP inbox messages with Jev category judgments, then tags, moves, flags, or notifies from a local Textual TUI and CLI. Project guide.
• Jev Mail Triage - Open source · Free source build · BYOK. Dry-run-first Proton Mail Bridge triage that uses TypeSafe Jev to flag spam and phishing before touching your mailbox. Project guide.
• Jev Moderation Bot - Source unverified · Pricing unverified · BYOK. Self-hosted Discord bot using Jev message classifications for automatic deletion, escalating timeouts and moderator-requested activity profiles. Project guide.
• Jev Personal Radar - Source unverified · Free source build · BYOK. Daily GitHub/HN radar: TypeSafe Jev ranks public candidates into private Issues; MiniMax explains winners (no LICENSE at tip). Project guide.
• Jev Playground by Laaaaksh (hosted bring-your-own-key playground for TypeSafe Jev) - Open source · Free · BYOK. Unofficial static web playground for trying Jev with your own TypeSafe key: paste a state (text or JSON), define questions, and see Jev's answers. A small relay function forwards your key to TypeSafe without logging it; you can also run your own copy. Project guide.
• Jev plays chess - Open source · Free · BYOK. Watch TypeSafe Jev play Maia-3 on an Elo ladder (facts-only Choice moves; GPL-2.0-or-later; deployed site AGPL via Maia). Project guide.
• Jev Quiz Pilot - Open source · Free source build · BYOK. Python CLI that lets TypeSafe Jev navigate web quizzes in Chrome and logs every pick for measurement. Project guide.
• Jev Radar - Open source · Free source build · BYOK. Local research workspace where Jev steers investigation over public sources with inspectable decisions and optional text-model drafting. Project guide.
• Jev Review (Gmail) - Open source · Free source build · BYOK. Chrome extension: TypeSafe Jev grades Gmail replies live while you write (distinct from other jev-review tools). Project guide.
• Jev Search - Open source · Pricing unverified · BYOK. Selects search sources and filters, then ranks retrieved links using Jev and Search1API. Try app · Project guide.
• Jev Sheets - Open source · Free source build · BYOK. OpenHarness desktop sheet pane: TypeSafe Jev answers typed column questions per row (Question Lab wording trials; BYOK or offline practice). Product · Project guide.
• Jev Situation Puzzle - Open source · Free · BYOK. Lateral-thinking puzzle game where TypeSafe Jev answers yes/no/irrelevant and judges story reconstructs (EN+ZH). Try app · Project guide.
• Jev Slop Detector - Source available · Free source build · BYOK. Unofficial Chrome extension (with a local Node relay holding your key) that labels posts on your X feed with a TypeSafe Jev classification probability such as ● Slop | 83%; it never hides or edits posts. The repo also holds SylphAI's class material for teaching Jev and the demo-recording scripts. Project guide.
• Jev Slop Guard - Open source · Free source build · BYOK. Chrome extension: TypeSafe Jev scores/stamps AI slop on X and LinkedIn feeds as you scroll. Project guide.
• Jev Social - Open source · Free source build · BYOK. Local UI/CLI where Jev, through OpenRouter or an explicit loopback System One-compatible provider, chooses read-only Instagram/TikTok/LinkedIn research ops executed by socai in Chrome. Project guide.
• Jev Songwriter - Open source · Free · BYOK. Code enumerates legal musical options; TypeSafe Jev picks mode/tempo/form/chords/notes with replayable demos. Try app · Project guide.
• Jev Soundboard - Open source · Free source build · BYOK. Live AI soundboard that listens to a call (FaceTime, Zoom, Meet, Discord or a stream), transcribes it and lets TypeSafe Jev decide whether to react and which clip to play within about a second; 15 built-in synthesized sounds, local control panel and an optional Discord bot. Project guide.
• Jev Studio (CarlosGuedea) - Source available · Free source build · BYOK. Spanish-language visual editor (React Flow + FastAPI) for building and running AI workflows where a Jev Decision node is the decision engine next to LLM, Python, HTTP, Condition and Output nodes; real backend execution with graph validation, decision branches and step logs, webhook / interval / cron triggers and SQLite history. Distinct from the listed Jev Studio CLI/MCP kit. Project guide.
• Jev Talks (chat built one Jev choice at a time) - Open source · Free. Chat experiment where a decision model writes the reply: on every step Jev picks the next character (English and Turkish letters, digits, punctuation, SPACE or STOP) or the next word from 250 common words, with an optional recheck pass. Project guide.
• Jev ToS Checker (Chrome extension that flags terms-of-service clauses that may work against you) - Open source · Free · BYOK. Chrome extension that, on request, splits the terms of service or privacy policy on the current page into clauses, has Jev check each against five risk types, and lists the flagged clauses with their original text, scrolling to each on the page; works on any site and in any language; English and Japanese UI. Project guide.
• Jev Town - Open source · Free source build · BYOK. Browser AI-town where TypeSafe Jev chooses each agent's structured actions and an LLM handles dialogue/memory. Project guide.
• Jev Trip - Open source · Free source build · BYOK. Local day-trip planner: TypeSafe Jev screens candidates and reviews while an LLM drafts; offline replay demo. Project guide.
• Jev Tweet Radar - Open source · Free source build · BYOK. Chrome (and iOS Safari Web Extension) scores each X post with one TypeSafe Jev call—engage/buzz/spam/AI-ish—and optional fade filters. Project guide.
• Jev Voice - Open source · Free source build · BYOK. Hands-free macOS voice assistant: local whisper.cpp plus one Jev call per command to select typed actions and arguments. Project guide.
• Jev Voice (CUA) - Source available · Free source build · BYOK. Native macOS floating bar: continuous voice/text; Jev picks live Accessibility actions in an observe–act–verify loop (no LLM planner). Unlicensed public source. Project guide.
• Jev Voice Chat (Japanese) - Open source · Free source build · BYOK. Japanese voice conversation in the browser where Jev handles turn-taking (end of utterance, backchannel vs interruption, restatement vs new content) and OpenAI models handle recognition, replies and speech. Project guide.
• Jev X Blocker (Chrome extension that hides high-confidence porn and spam on X with your own Jev key) - Open source · Free · BYOK. Local-first Chrome extension that uses your own TypeSafe key to cover high-confidence pornographic, sexually suggestive and porn-funnel posts and replies on X with restorable placeholders, across timelines, search, profiles and threads; a refactor and extension of bugkiwi's elons-job; Chinese documentation. Project guide.
• Jev X Sentiment Analysis (jev-v8) - Open source · Free source build · BYOK. Self-hosted crypto decision-support terminal: statistics pre-process tweets and market data, TypeSafe Jev answers trade/sentiment/squeeze/catalyst questions; no auto-trading. Project guide.
• Jev · Bridge by Bridgehead (hosted MCP connector that puts Jev's choices, scores and yes/no checks in Claude, ChatGPT and other MCP apps) - Closed source · Pricing unverified · BYOK. Hosted MCP connector that brings Jev's typed questions, ranking, file questions, an agent gate and people search to Claude, ChatGPT and other MCP apps on your own key (vendor description). Try app · Project guide.
• Jev × WebMCP - Open source · Free source build · BYOK. Chrome side panel that selects and populates page WebMCP tool calls with TypeSafe Jev from keystrokes (Chrome 149+ / WebMCP; BYOK; load unpacked). Project guide.
• Jev 棋馆 (Jev board games) - Open source · Free source build · BYOK. Browser board-game hall (Gomoku, Go 9x9, Xiangqi, chess, checkers, Chinese checkers) where TypeSafe Jev chooses moves, with a deterministic safety layer and a Cloudflare Worker + D1 record book. Project guide.
• jev-crypto-scout - Open source · Free source build · BYOK. Crypto screening: code owns CoinGecko signals; TypeSafe Jev judges news (not a trading bot). Project guide.
• jev-demofast - Open source · Free source build · BYOK. One sentence in, a narrated product demo out: TypeSafe Jev drives your real product in a browser (optional source-built screen map). Try demos · Project guide.
• jev-desktop - Open source · Free source build · BYOK. Windows desktop agent: UI Automation observes controls, TypeSafe Jev chooses the next action, and a small LLM types text when needed (jevd). Project guide.
• jev-email-classifier - Open source · Free source build · BYOK. Chrome MV3 extension that classifies Gmail rows with TypeSafe Jev (Choice/Score/Noul); metadata-only, BYOK. Project guide.
• jev-ercot - Open source · Free source build · BYOK. Texas retail electric plan shopper over the Power to Choose corpus: TypeSafe Jev classifies Electricity Facts Label spans; Next.js ranks plans by projected 12-month cost. Project guide.
• jev-face (Imjungjuna) - Open source · Free source build · BYOK. WebGL 3D face whose ARKit blendshapes react in real time as TypeSafe Jev judges the emotion of what you type (distinct from rbrisuda/jevface Java client). Project guide.
• jev-gmail-filter - Open source · Free source build · BYOK. Local Gmail topic filter UI: plain-English topics judged via jevfilter + TypeSafe Jev; OAuth and data stay on your machine. Project guide.
• jev-leads - Open source · Free source build · BYOK. Inbound lead qualification via OpenRouter Jev-1.13 with SQLite zero-loss buffer and Airtable CRM sync (MIT; distinct from Airtale). Project guide.
• jev-mail (vynnlee) - Open source · Free source build · BYOK. Autonomous Gmail zero-inbox triage powered by TypeSafe Jev System One (config-driven). Project guide.
• jev-minecraft - Open source · Free · BYOK. Runnable demo (Chinese docs) in an open-source Minecraft-compatible world: each step sends the bot's position, inventory, nearby blocks, task progress and available options to TypeSafe Jev, Jev picks the next action, Mineflayer executes it and the real world is checked; tasks range from collecting logs and building a pillar to a 90-block camp, and you can join the same world from a browser player page. Project guide.
• Jev-Mod (undeemed) - Open source · Free source build · BYOK. Discord moderation bot powered by TypeSafe Jev rules: hosted dashboard or self-host on Cloudflare/Docker with your own API key. Project guide.
• jev-x-filter - Open source · Free source build · BYOK. Chrome MV3: TypeSafe Jev filters X timeline spam with dry-run mute/block defaults. Project guide.
• Jev/ui (jev-genui) - Source available · Free source build · BYOK. Generative UI sandbox: TypeSafe Jev walks a UI grammar; React renders live. Project guide.
• Jevassic Park - Open source · Free. A shared, live isometric theme park where about 100 individually simulated guests pick rides, snacks and spectacles through Jev while code handles movement, queues and rides; anyone can watch at the hosted site. Project guide.
• Jevatar - Open source · Free source build · BYOK. Local face-only companion: TypeSafe Jev chooses one of 16 moods; blobatar morphs the face (no text replies). Project guide.
• jevbar - Open source · Free source build · BYOK. NL query box that composes typed UI screens (forms/confirmations/rankings) via TypeSafe Jev. Project guide.
• JevBystander - Open source · Free source build · BYOK. Android Accessibility WeChat reader: TypeSafe Jev Toast judgments only (no reply text/send). Project guide.
• Jevcast - Open source · Free source build · BYOK. Native macOS launcher/window manager with optional TypeSafe Jev natural-language action matching. Product · Project guide.
• JevChess - Open source · Free source build · BYOK. Mobile-friendly 3D browser chess where TypeSafe Jev chooses among legal moves instead of writing move text. Project guide.
• Jevchik (Telegram group bot: karma for useful messages, anti-spam for newcomers, judged by Jev) - Open source · Free source build · BYOK. Self-hosted Telegram bot for group chats: members earn karma from reactions, replies, thanks and from messages Jev judges useful, a karma number appears next to their names, newcomers get spam screening, and a Mini App shows weekly, monthly and all-time leaderboards. Bot texts are in Russian; docs in Russian and English. Project guide.
• JevClean - Open source · Free source build · BYOK. Local IMAP mail organizer: write natural-language rules (one action each), Jev judges every message against them, and you review and confirm delete, archive, move, spam or flag actions with undo. Project guide.
• JevDash - Open source · Free source build · BYOK. Clean-room 2D platformer bench for TypeSafe Jev real-time control at 60 FPS. Project guide.
• JevDK and jev-serve (Mac tools to design decision questions and serve a Jev-style API locally) - Open source · Free. Two free Mac tools: JevDK lets you design and test typed decision questions on a local MLX model and on hosted backends side by side (including TypeSafe Jev through OpenRouter), mark right answers and calibrate; jev-serve answers a hosted-Jev-style HTTP API on your Mac. Project guide.
• jevelry - Open source · Free source build · BYOK. Terminal app to make and track decisions with TypeSafe Jev (jevels). Project guide.
• JevEye - Open source · Free source build · BYOK. Browser vision probes report calibrated facts; TypeSafe Jev plans probes and judges the text fact sheet (never pixels). Project guide.
• Jevfill - Open source · Free source build · BYOK. Chrome extension that autofills forms from unstructured notes with TypeSafe Jev field matching (skips password/payment). Distinct from Smart Paste. Project guide.
• Jevflix - Open source · Free source build · BYOK. Hybrid FAISS+BM25 movie shortlist; TypeSafe Jev parses constraints and picks one film with a confidence gate. Project guide.
• JevGuide - Open source · Free source build · BYOK. Android WeChat relationship assistant: TypeSafe Jev judges affinity/intent; chat model drafts three overlay replies (never auto-sends). Product page · Project guide.
• JevIntent - Open source · Free source build · BYOK. FkWeChat WeChat plugin: TypeSafe Jev judges intent/emotion/stance on long-press (no reply text/send). Project guide.
• JEVision (flowchart UI tests for any Mac app, with Jev checking screens and finding elements) - Open source · Free source build · BYOK. macOS app for visual UI testing of any Mac app, including ones you cannot instrument: a test is a JSON flowchart walked block by block with the real mouse and keyboard, and Jev can check what a screenshot shows with a confidence threshold and find elements described in words such as "the blue Save button". Project guide.
• Jevline - Source available · Free · BYOK. Incident-timeline proof of concept for security analysts: start from one confirmed-malicious process and Jev scores, process by process, which other activity belongs to the same incident, producing a reviewable timeline (web portal, local portal or CLI). Project guide.
• Jevmail - Open source · Free source build · BYOK. Local read-only Gmail triage into Needs reply/Updates/Promos/Sales/Spam with TypeSafe Jev over the Vercel AI Gateway. Project guide.
• Jevmail (hwong103) - Open source · Free source build · BYOK. Self-hostable read-only Gmail triage where TypeSafe Jev sorts mail into trays you can correct (Cloudflare Worker + D1 + React PWA). Project guide.
• JevMail (Outlook) - Open source · Free source build · BYOK. Outlook add-in for Microsoft 365 work and school accounts, powered by TypeSafe Jev. Project guide.
• jevmdb - Open source · Free source build · BYOK. Find similar movies with TypeSafe Jev over IMDb/TMDB/Wikidata metadata and where-to-watch links. Project guide.
• Jevmeter - Open source · Free source build · BYOK. Renders video captions, gauges and highlights from Jev judgments about transcript sentences; these judgments are not fact-checks. Project guide.
• jevmoji (web app that suggests emoji reactions for a chat message by scoring 200 emojis with Jev) - Open source · Free source build · BYOK. Small Go web app: type a line of group-chat banter and the server scores all 200 catalog emojis in one Jev call, one yes/no question per emoji, and shows the best four (50% or higher) plus a fifth above 75% as Slack-style reactions. HTMX front end embedded in a single static binary; Docker and Cloud Run deploy steps included. Built as an example integration of the Jev API. Project guide.
• JevPaste - Open source · Free source build · BYOK. macOS menu bar Smart Paste: TypeSafe Jev selects an exact clipboard/profile value for the focused field. Project guide.
• JevPDF - Open source · Free · BYOK. Ask a PDF in your own words; TypeSafe Jev answers one yes/no question per extracted line and matching lines highlight in the browser. Try app · Project guide.
• JevRadar (local-first Windows desktop radar for market sentiment and context, classified with Jev) - Open source · Free source build · BYOK. Tauri desktop app (Portuguese UI, bilingual setup) that combines public content from RSS, Reddit and optional X with Jev's typed classifications and real market series, to show shifts in attention and measure what happened after past signals. It does not recommend trades or execute orders. Project guide.
• Jevry - Open source · Free source build · BYOK. Desktop browser agent: TypeSafe Jev chooses the next executable page action; a separate text/vision model plans. Project guide.
• JevSlop - Open source · Free source build · BYOK. Scores note articles for AI-slop writing patterns with TypeSafe Jev (BYOK). Try app · Project guide.
• JEVsnap - Open source · Free source build · BYOK. Screenshot hotkey that blacks out secrets and personal information (API keys, tokens, emails, phone numbers, names, internal hostnames, QR codes) before the image reaches your clipboard; on-device OCR and local rules first, with optional Jev text and Box checks. Project guide.
• JevTRPG - Open source · Free · BYOK. Paste a résumé or backstory and get a 1920s TRPG investigator sheet judged by TypeSafe Jev (free live site; nothing stored). Try app · Project guide.
• jevx - Open source · Free source build · BYOK. Chrome/Firefox extension: find relevant X posts and score drafts with TypeSafe Jev (BYOK; no backend). Project guide.
• JevZero - Open source · Free source build · BYOK. Local Gmail triage with TypeSafe Jev: review proposed labels, apply with receipts, and undo. Project guide.
• Jev会計 (jev-kaikei-free) - Open source · Free source build · BYOK. Local Japanese bookkeeping tool: TypeSafe Jev picks account, tax class and private-use flag per receipt; low-confidence items go to human review; mock mode without a key. Project guide.
• jizhang (self-hosted subscription and renewal tracker with optional Jev categories) - Open source · Free source build · BYOK. Self-hosted command center for recurring expenses: subscriptions, renewals, reminders and the decisions around them; optional per-user Jev classification picks a category when a subscription is added. Project guide.
• job-drill - Open source · Free source build · BYOK. Local workplace response drill: write timed replies to scenario prompts and get Jev scoring reports. Project guide.
• Keel - Open source · Free source build · BYOK. Local-first macOS coding workspace: local Laya selector by default, optional hosted TypeSafe Jev route selection. Project guide.
• keyed james bible - Open source · Free source build · BYOK. Turns a question into a span of King James Version verses by keying it through TypeSafe Jev choice requests (book and extent → chapter → verse) and clipping the passage verbatim from a local KJV corpus; nothing is generated, and the author documents measured Jev quirks. Project guide.
• KeyFlip (Windows tray app that picks the right Hebrew or English layout per text field) - Open source · Free source build · BYOK. Windows tray app for people who type in Hebrew and English: when you focus a text field it chooses the layout from what it learned for that field, the text already there, or Jev's reading of the app, site, page title and field label, and it fixes a word you typed on the wrong layout. Project guide.
• kureksistant - Source available · Free source build · BYOK. Arch Linux (Hyprland) personal AI assistant that uses TypeSafe Jev for sub-second decisions. Project guide.
• Kuro (Chrome reading aid for ADHD with bold word starts, a focus reader and optional Jev paragraph triage) - Open source · Free source build · BYOK. Manifest V3 Chrome extension that bolds the start of every word, enlarges the paragraph you are reading while the rest of the page dims, and optionally summarizes dense pages. Jev scores how dense each paragraph is and whether it carries the argument, so only the right paragraphs are turned into bullets or marked in the margin. Project guide.
• L Ragebait - Open source · Free. Chrome extension that conservatively hides high-confidence ragebait, hostile threads and sensationalised tragedy commentary on X and YouTube, using three Jev noul scores per item in one request against a threshold you set (default 0.85). Project guide.
• Laser (Omarchy focus mode) - Open source · Free · BYOK. AI focus mode for Omarchy / Hyprland on Linux: you declare a task, and Laser judges each screen against it with TypeSafe Jev (about half a second per check), escalating from a red dot to a nudge, a red screen edge and a haze over the distracting window; screenshots are OCR'd on device and only text (per privacy level) is sent. Project guide.
• LifePot - Open source · Free source build · BYOK. Replayable artificial-life game where Jev proposes species roles, relationships and bounded rule changes from your three answers, and validated rules (not code) drive a deterministic ecosystem. Project guide.
• Liminal Drift - Open source · Free source build · BYOK. First-person Backrooms horror prototype with TypeSafe Jev realtime scene decisions. Project guide.
• LingYi App Watch (local macOS app-update dashboard, with optional Jev upgrade-risk ratings) - Open source · Free. Local-first macOS update dashboard that gathers pending updates from Homebrew, the Mac App Store, GitHub Releases, Sparkle appcasts and other sources, flags third-party sources and lets you mark apps as do-not-upgrade or to-verify; with a TypeSafe key, each pending update also gets a Jev risk level, confidence and reason. Chinese-language UI and docs. Project guide.
• LinkedIn Slop Filter - Open source · Free source build · BYOK. Chrome extension + local proxy: TypeSafe Jev stamps LinkedIn posts Bait/Corp/Brag/Slop (distinct from Slop Mop). Project guide.
• LinkScout - Open source · Free source build · BYOK. Browser extension plus your own Cloudflare Worker that scores Google, Bing and DuckDuckGo results before you click: pages are fetched as Markdown and TypeSafe Jev judges relevance, depth, SEO spam and category and picks a key passage. Project guide.
• Liuyao - Source available · Free source build · BYOK. Chinese-style I Ching (六爻) site with programmatic casting plus TypeSafe Jev semantic judgment (sixyao.app; non-commercial source). Project guide.
• Live Jev - Open source · Free source build · BYOK. Early macOS Ableton Live controller: hotkey bar + TypeSafe Jev picks typed mixer/device actions from short EN/JP phrases. Project guide.
• Llasso (draw a box round anything on your Mac and ask about it or have it changed) - Open source · Free · BYOK. Mac app where you hold ⌘ and drag a box round anything on screen, then type or say what you want: Llasso guesses what the box holds, picks a model for the kind of work from OpenRouter's live lists, and every change waits for your tap; Jev on OpenRouter classifies the request when it can answer within 1.5 seconds. Project guide.
• LYRCENSOR (C terminal app that redacts slurs in Russian and English lyrics using Jev probabilities, editing letters locally) - Open source · Free source build · BYOK. Native C17 terminal application for context-sensitive lyrics redaction: it asks one OpenRouter Decisions model (Jev by default) for noul probabilities and replaces selected letters locally in C, so the model never writes output text. Russian and English text, line endings and spacing are preserved. Project guide.
• Mac Voice Control - Open source · Free source build · BYOK. Hands-free macOS control by voice: Deepgram or local Whisper speech-to-text, typed action routing by Jev (typesafe/jev through Command Code's provider API) or an LLM over a closed action set, and AppleScript/Accessibility execution; open apps, drive browser tabs and dictate code into IntelliJ or VS Code. Project guide.
• macbrow - Open source · Free source build · BYOK. Experimental macOS voice assistant using Jev to route commands to AppleScript tools and Chrome browser tasks. Project guide.
• Magic Jev Ball - Open source · Free. Magic 8 Ball in the browser that doesn't pick at random: hold it, shake it, let go, and Jev picks which of the 20 classic answers fits your question, with a probability for every answer shown under the ball. Project guide.
• mail-watcher - Open source · Free source build · BYOK. Self-hosted Cloudflare Email Worker that redacts incoming mail, lets Jev pick a per-watch category, and forwards, drops, or notifies accordingly. Project guide.
• Mailroom - Open source · Free source build · BYOK. Gmail sorter with readable rules plus TypeSafe Jev typed judgments, receipts, and undo. Project guide.
• Mardi Gras - Open source · Free source build · BYOK. Terminal UI for Beads agent issues that can use Jev as an optional fast judge for duplicate checks, focus order, formula choice and Codex approval risk advice. Project guide.
• Masroufi - Source available · Free source build · BYOK. Gaza household expense tracker: TypeSafe Jev classifies statement notes with confidence and human review. Project guide.
• meeting-ledger - Open source · Free source build · BYOK. Terminal app that keeps meetings on the record: Claude extracts decisions, action items, promises and open questions with cited transcript lines, a System One model (TypeSafe Jev or local Kev) checks proposed status changes across meetings, and anything the two disagree on waits for your review; stale promises are flagged. Project guide.
• Meraline - Open source · Free · BYOK. Mac ⌥Space assistant for any app with a Decision mode: ask yes/no, choice or level questions about selected text and get Jev (direct or via OpenRouter) or local decision-model answers with how sure they are. Project guide.
• Metis Suite (self-hosted mail and calendar client with an optional Jev decision model) - Open source · Free source build · BYOK. Self-hosted mail, calendar and drive client with an AI agent and workflows for Gmail, Outlook, Zoho or any IMAP/SMTP and CalDAV server; an optional decision model (TypeSafe's Jev by default for the systemone driver) labels inbox mail, gates thread summaries and powers workflow Ask and Branch steps. Project guide.
• Mimir (self-hosted news reader that ranks RSS feeds against your interests and groups stories with Jev) - Open source · Free source build · BYOK. Self-hosted news reader: collect any number of RSS and Atom feeds, describe your interests in plain words, and Mimir scores every article against them and groups articles about the same event into one story with a timeline. Jev does the scoring and story matching; on-demand sourced briefs are written by a MiniMax model. Project guide.
• MinusPod (self-hosted podcast ad remover with native System One support) - Open source · Free source build. Self-hosted server that transcribes podcast episodes, detects and cuts ad segments, and learns sponsor patterns from your corrections; it has a native System One adapter so Jev can run ad detection and verification. Project guide.
• Mish - Open source · Free source build · BYOK. Local-first AI design workspace where TypeSafe Jev routes briefs and verifies facts before publish (agency fast lane). Project guide.
• Mobile Agent (KYRIE66nb) - Open source · Free · BYOK. Open-source on-device Android AI agent (Chinese/English) that sees and taps the screen, with an optional dedicated decision backend — TypeSafe Jev or self-hosted Laya over /v1/systemone — for safety-gate verdicts and low-risk navigation choices; off by default, SHADOW/ENFORCE modes, explicit consent. Project guide.
• Model vs Market - Open source · Free · BYOK. Web app where three decision models — OpenAI Decisions, TypeSafe Jev and Cloudflare Clef — read recent news (never the odds), put a probability on Polymarket and Kalshi questions, and are compared with live market prices, with a replay of how each model changed its mind article by article. Project guide.
• Modex - Open source · Free source build · BYOK. TypeSafeAI open Codex-style desktop agent; optional Jev Auto routing for model/effort. Project guide.
• Moji (macOS emoji picker with optional Jev ranking) - Open source · Free source build. Native macOS emoji picker: type : and a few letters in any app. Search runs locally; with a TypeSafe key, Jev reorders the suggestions to the emoji you most likely mean. Project guide.
• nameworthy (World of Warcraft: Forever name auditor that grades a name with one Jev call) - Open source · Free source build · BYOK. Single-binary Go web app for fun: enter a WoW: Forever first name and surname, optionally pick race and class, and get a grade from S to F, a naming style, a six-stat sheet and probability signals such as 'likely to be reported', all from one System One call with the full distribution shown. Project guide.
• Nature Lens - Open source · Free source build · BYOK. Upload a nature photo and Jev ranks which landscape categories are visible, picks the main subject and draws a box around any person (Jev 1.14 Box), all in one request. Project guide.
• Nectra - Open source · Free source build · BYOK. One chat across Claude Code, Codex, Gemini CLI and Ollama; TypeSafe Jev classifies each message for routing. Project guide.
• Neon Chase - Open source · Free source build · BYOK. Neon-style endless runner where TypeSafe Jev can drive realtime gameplay decisions. Project guide.
• Netizen (simulate how Indonesian netizens react to a post before it goes live, judged by Jev) - Open source · Free · BYOK. Simulator by Moddest Lab: retrieve similar past cases and their real comment mix, let Jev judge which personas (Bekasi mothers, ojol drivers, students, buzzers, news accounts and more) see the post, then run two rounds of reactions across niche, busy and viral reach scenarios. Text and image posts; a hosted demo where you bring your own key. Project guide.
• Neutron Email Triage - Open source · Free source build · BYOK. Jev-first email classifier for Proton Mail through Proton Mail Bridge, with a Windows Electron GUI and a source CLI: dry-run by default, with optional moves to trash or archive above a confidence threshold. Project guide.
• Noa - Open source · Pricing unverified · BYOK. Chrome automation extension where Claude Code plans and TypeSafe Jev picks each click and keystroke. Project guide.
• NotForMe (Chrome extension that folds PR, product-launch and off-topic articles in Korean news lists using Jev) - Open source · Free · BYOK. Chrome extension (README in Korean and English) that collapses, blurs, dims or hides advertorial PR articles, product-launch pieces and off-topic stories in the article lists of ZDNet Korea, Yonhap, Hankyoreh and other Korean news sites. It only changes how the list is shown in your browser. Project guide.
• Notiq - Source available · Free source build · BYOK. Android notification filter: natural-language rules + TypeSafe Jev/FastJev keep/filter Choice (no LICENSE file). Project guide.
• NotiSift (Android notification organizer where a decision model such as Jev sorts each notification) - Open source · Free. Open-source Android app that rebuilds your notification drawer: each notification is decided as Important, Later or Junk and grouped, with important ones on top and the rest in quiet day and group stacks. Decision models (Jev, Perplexity's, or a self-hosted one) or chat models can do the sorting with your own key; codes, chats, calls and alarms never leave the phone, and your own rules always win. Works offline without AI. Project guide.
• Notra - Open source · Commercial · Paid. Uses Jev judgments within a broader application for tracking brand mentions and placement in AI answers. Product · Pricing · Project guide.
• Nudge for Omarchy - Open source · Free source build · BYOK. Omarchy (Hyprland) launcher plugin: describe what you need in plain words and Jev routes it to a reviewed safe desktop operation (night light, brightness, volume, folders, settings panels, reminders, shortcut lookup) or shows a preview and manual guidance for anything else. Project guide.
• Omarchy Mail - Open source · Free source build · BYOK. Terminal Gmail client for Omarchy; TypeSafe Jev files inbox into labels. Project guide.
• omarchy-mouse-sacrifice - Open source · Free source build · BYOK. Omarchy (Hyprland) plugin: move the pointer into the bottom-right corner and a card offers up to five keyboard shortcuts that fit what is on screen, ranked by TypeSafe Jev from a list built in code (focused app, workspace windows, bar widgets, recent use, your real Hyprland binds and press counts); no screenshots or typed text are sent. Project guide.
• open-annie - Open source · Free source build · BYOK. Voice-driven 3D character in the browser: GPT-Live-1 talks with you over WebRTC while TypeSafe Jev reads the live transcript and picks every face, gesture and dance (51 decisions at p50 118 ms and $0.0026 of Jev in the recorded session); lip-sync and rendering run in the browser behind a thin broker on fal serverless. Project guide.
• OpenMausBot (desktop chat app for a team of AI agent bots, with Jev picking which bot answers) - Open source · Free · BYOK. Electron desktop app for macOS, Windows and Ubuntu where each bot in a chat-style sidebar is a local Claude Code, Codex or Grok CLI agent with its own personality, model, computer and connected apps. With a decision model enabled, rooms set to Auto send each un-mentioned message to Jev, which picks the bot that should answer in a few hundred milliseconds; the room's lead bot answers whenever the decision model is off, slow or unsure. Project guide.
• Orbs - Open source · Free source build · BYOK. Self-hosted multi-bot chat: rooms hold people and bots, @mentions wake a bot, and with no mention TypeSafe Jev scores which bot should answer; a daemon on your machine runs each bot turn through your own Pi agent. Project guide.
• Orphaned Films - Open source · Free. Free movie browser and player for public-domain films on archive.org (orphanedfilms.com) where Jev decided offline which TMDB film each upload really is and which films lead each collection shelf. Project guide.
• PageGrade - Open source · Free source build · BYOK. Chrome extension that grades page sections for clarity, writing, and on-page SEO with TypeSafe Jev via Vercel AI Gateway. Project guide.
• Paper Radar - Open source · Free source build · BYOK. Daily arXiv radar: TypeSafe Jev judges new papers against plain-English interests; GitHub Pages + RSS digests. Project guide.
• PaperRadar - Open source · Free source build · BYOK. Browse and tag conference papers locally; TypeSafe Jev applies research tags, export to Zotero. Project guide.
• PaperWing (open-web reader organised by topic, with Jev grading how heated a discussion is) - Open source · Free. Ad-free reader that pulls Hacker News, Lobsters, Bluesky, RSS, arXiv and engineering blogs into topics rather than sources, merges duplicate links and ranks across sources with a recipe that runs on your device. Jev powers an optional spice rating: a score question and four yes/no questions grade how heated a thread's discussion is. Project guide.
• Pastewise - Source unverified · Free · BYOK. One paste box morphs into JSON/JWT/cron/stack/color tools; TypeSafe Jev classifies ambiguous pastes (offline heuristic by default; live demo). Try app · Project guide.
• patrol-jev - Open source · Free source build · BYOK. Photo-to-patrol-log web app: OpenAI describes photos; TypeSafe Jev classifies the write-up. Project guide.
• Penny (grocery receipt review for Grocy, with Jev suggesting storage and groups) - Open source · Free source build. Self-hosted receipt review for Grocy: upload a grocery receipt PDF, OpenAI extracts the lines, Jev suggests a storage location and product group for each product (low-confidence ones stay unassigned), and confirmed purchases are imported into Grocy stock. Project guide.
• PetPaw - Source available · Free · BYOK. Desktop pet companion for Apple Silicon Macs (SwiftUI, SceneKit, AVAudioEngine) with imported characters, voice and lip-sync: interaction reactions and poses are chosen by Cloud JEV (typesafe-ai/jev) or a Local Open-Jev mode that runs a 2B Open-Jev checkpoint ported to Swift/MLX on device, while Apple Foundation Models write moods and speech bubbles. Project guide.
• Photo Ferret - Open source · Free. Self-hosted photo gallery where Jev's image questions rate every photo on dozens of scales (how much sky, water or snow, sky drama, wildness, close-up or vista, quality), so visitors filter a static Cloudflare site with sliders, chips, a colour picker and an experimental 3D map. Project guide.
• pitwall - Open source · Free · BYOK. Native terminal multiplexer for coding agents (live status, persistent sessions, tabs/panes) that can connect TypeSafe Jev (pitwall jev login) to answer quick questions about your agents: is this permission request safe, how urgent is this pane, has a hook-less agent finished, does this turn need review. Project guide.
• Planet.buzz (Reddit-style discovery feed from many web sources, with Jev classifying topics, motivations and content warnings) - Closed source · Pricing unverified. Old-Reddit-style discovery feed from many web sources where Jev classifies topics, motivations and content warnings and app code ranks posts (maker claim; self-submitted). Try app · Project guide.
• Plasis - Open source · Free source build · BYOK. One text box that turns what you type into the right UI card (event, reminder, issue, checklist, bill split and more): one TypeSafe Jev call answers 15 typed intent questions in parallel while a deterministic parser reads dates and amounts; a free offline keyword classifier is the default, with measured agreement against Jev. Project guide.
• Please.exe - Open source · Free. Private tone coach for Claude/ChatGPT drafts with an optional, confirmed TypeSafe Jev second opinion. Project guide.
• Pointsman by RomainJeff (macOS link router that picks the right browser with Jev) - Open source · Free source build · BYOK. macOS menu bar app you set as the default browser: describe once how you use each real browser, and each clicked link goes to the right one, chosen by Jev from your descriptions, the app the link came from and the URL without its query string. Project guide.
• Polymorph - Open source · Free source build · BYOK. Chrome extension that collapses posts matching English rules judged by TypeSafe Jev via OpenRouter Decisions; replace with your media or restore. Project guide.
• PR Review (sunwrobert) - Open source · Free source build · BYOK. Keyboard-first macOS app for reviewing and merging GitHub pull requests through your gh login, with optional Jev readiness scores, smart search and grouping of related PRs. Project guide.
• Preguntale a Jev - Source unverified · Free · BYOK. No-login yes/no chat: TypeSafe Jev Choice probabilities via Vercel AI Gateway (Spanish UI). Try app · Project guide.
• pulp - Open source · Free source build · BYOK. Personal macOS pipeline that turns Safari iCloud tabs into clean EPUBs for an e-ink reader (OPDS catalog, KOReader sync, optional reMarkable push); optional --smart triage asks TypeSafe Jev whether each tab the denylist kept is worth reading. Project guide.
• Qiaomu Banbot (乔木净评) - Free · Source available · BYOK. Chrome extension preview that uses TypeSafe Jev (or Magpie / OpenRouter / Vercel gateways) to fold suspected AI, low-signal, and promo replies on X, with optional natural-language rules and hot-post probability estimates. Project guide.
• Qoresence - Open source · Free · BYOK. Local-first Windows 'gaming streaming observatory' for PS5 + capture card + DualSense (Madden 27 / College Football): HDMI video, controller HID and game situation become one causal event bus surfaced through overlays, clips and a pull-only AgentGlass/MCP; with opt-in TypeSafe Jev, typed score-plausibility and ticket checks keep the board dark instead of painting an unconfirmed score — Jev never mints score digits. Project guide.
• QuantDinger - Open source · Free source build · BYOK. Self-hosted AI trading OS with optional TypeSafe Jev pre-trade entry gates; hosted app also available. Product · Try app · Project guide.
• QuickAsk (macOS menu bar assistant for Codex and Claude, with Jev routing and follow-up checks) - Source available · Free source build · BYOK. Native macOS menu bar assistant: press Option–Space, ask a question and send it to Codex or Claude using your existing CLI logins; Jev (or optionally OpenAI Decisions) routes which context to attach and checks whether an answer needs a follow-up. Project guide.
• Quiet X - Open source · Free source build · BYOK. Open-source Chrome extension that silently hides X ads and posts by account country, language or topic; optional language/topic rules are judged by TypeSafe Jev on the rendered post text through a small local proxy that holds your key. Project guide.
• QuietFlow (Android notification attention manager with opt-in Jev classification) - Open source · Free source build · BYOK. Android proof of concept that listens to notifications and sorts them into five attention classes, from action required to hide; local safety rules protect OTP, security, finance, calls and system notifications, Jev classifies the rest when you opt in, and observation mode (never dismissing anything) is the default. Project guide.
• Railroad Route - Open source · Free source build · BYOK. Mine-cart track puzzle: lay and turn track pieces, then write the cart a one-sentence note; TypeSafe Jev-operated switches read the sentence and decide which way to send it, and each level forbids the obvious words, so you must describe the goal indirectly (English or Portuguese). Project guide.
• Recall (Chrome history search with a local index, browsing stats and optional Jev ranking) - Open source · Free · BYOK. Desktop Chrome extension for finding a page you remember visiting. It builds a local index of your history, shows a browsing overview, and if you opt in with your own TypeSafe key, Jev re-ranks a shortlist of local matches by meaning. Project guide.
• Reddit Pain - Open source · Free · BYOK. Chrome extension: TypeSafe Jev highlights Reddit posts asking for a product or describing a pain (BYOK). Project guide.
• RedOS - Open source · Free source build · BYOK. macOS menu-bar assistant for voice and text commands where a System One model using the Jev protocol routes request kinds and decides whether to speak or stay quiet. Project guide.
• RefGarden - Open source · Free source build · BYOK. Local spatial reference gallery where TypeSafe Jev chooses search phrases and highlights Met/NASA/Cosmos/Archive items from text metadata. Project guide.
• Refix - Closed source · Paid · Commercial. Growth: AI that helps your product grow faster on autopilot by running product experiments, SEO, content, and ads. Project guide.
• Reflex (monotykamary) - Open source · Free source build · BYOK. Mac menu-bar app to see, answer and talk to every coding agent running in herdr; Jev (via jev-fabric) routes each request to the right session or folder and asks you to confirm when unsure. Project guide.
• Regret Check - Open source · Free source build · BYOK. Chrome extension that pauses regret-prone clicks scored by TypeSafe Jev via a local OpenRouter-backed server. Project guide.
• Resenha (free push-to-talk dictation for macOS with local Whisper and optional Jev corrections) - Open source · Free · BYOK. Native push-to-talk dictation for Portuguese, English and Spanish: hold a shortcut, speak, and whisper.cpp transcribes on your Mac and types into the active field. Optionally, Jev judges only the transcript and local alternatives to pick contextual corrections; when it is unsure, the original text wins. A public unsigned Windows beta also exists. Project guide.
• Resurface - Open source · Free source build · BYOK. Resume screening: LLM writes questions; TypeSafe Jev System One classifies answers. Project guide.
• RimWorld Autopilot - Open source · Free. Experimental autonomous colony manager for RimWorld 1.6 on Windows: Laya — the open, Jev-compatible System One decision model — runs locally, reads the map through a bundled RIMAPI mod, picks what matters next and issues real in-game orders (food, building, work, trade, caravans, raids), with a desktop app to watch its reasoning and change priorities. Project guide.
• River Oaks - Source available · Free source build · BYOK. Interactive NPCs of River Oaks, Houston, with TypeSafe Jev powering NPC decisions. Project guide.
• say-hi - Open source · Free source build · BYOK. Fun experiment: TypeSafe Jev types chat replies one keystroke at a time (chooser-as-chat). Project guide.
• ScamCheck - Open source · Free source build · BYOK. Paste a suspicious message and get a scam verdict and risk score: one TypeSafe Jev call answers about eleven typed questions, and code can only raise the risk. Project guide.
• Scene Finder - Free · Source available · BYOK. Desktop app for finding a moment in a movie you can only describe: it samples frames from your MKV or MP4, asks Jev which ones match, lets you play the hits, then trim and save an MP4 clip. Project guide.
• ScrollPatrol - Open source · Free source build · BYOK. Chrome extension that mutes feed posts by meaning with TypeSafe Jev Noul rules (LinkedIn/Reddit/HN + short-video beta). Project guide.
• SearchJev - Open source · Free · BYOK. Firefox extension (Japanese UI) that asks Jev six questions about each Google result, then shows 0 to 100 scores and reorders the page to push mass-produced SEO and affiliate articles down and first-hand sources up; one click restores Google's order. Project guide.
• SeeFood (food recognizer with EmbeddingGemma and Jev) - Source available · Free. Web app inspired by the hot dog app from Silicon Valley that recognizes many dishes: EmbeddingGemma 2 retrieves candidate foods from mean-centered visual prototypes, and Jev image picks among them. Project guide.
• seo-jev-analyzer - Open source · Free source build · BYOK. Measures how closely each page matches its search intent using TypeSafe Jev with a calibrated noise band. Project guide.
• Shapes of Stories (story arcs of books and scripts scored with Jev) - Open source · Free source build. Charts the shape of a novel, screenplay or transcript: Jev scores every passage of about 800 words on ten qualities such as fortune and tension plus events, and the fortune curve is matched to named arcs like Rags to Riches or Man in a Hole. Project guide.
• Shapeshift - Open source · Free · BYOK. One text box morphs into the right UI via TypeSafe Jev intent (offline keyword classifier by default; live demo). Try app · Project guide.
• Should I Jev? - Open source · Free · BYOK. Describe a feature; get a verdict on code vs TypeSafe Jev vs LLM/ML (nineteen parallel Jev questions). Try app · Project guide.
• Sift (Chrome extension that reranks Google results with Jev and folds away sales pages and SEO filler) - Open source · Free source build · BYOK. Chrome extension that moves Google results that actually answer your query to the top and folds away sales pages and SEO filler. Each search gets one extra question about whether it is transactional ("buy nike pegasus 41"), so shopping searches keep their sales pages. Project guide.
• Signal & Noise - Source unverified · Free source build · BYOK. Chrome extension that hides X posts matching plain-language mutes (rage bait, crypto shilling, spoilers, your own) by asking TypeSafe Jev one choice question per tweet before it scrolls into view, with a why-bar and local feedback log. Project guide.
• Signal Feed (X and YouTube bait filter with Jev) - Open source · Free source build. Chromium extension that folds ragebait, empty hype and engagement bait on the X timeline and YouTube before you scroll to them, each with the reason, highlights posts on topics you choose, and shows what the Jev calls cost. Project guide.
• SignalLens - Open source · Free source build · BYOK. Local-first chat signal analyzer (Windows portable): redacts on-device, then TypeSafe Jev judges emotion/intent/engagement. Project guide.
• Simple Launcher - Open source · Free source build · BYOK. Minimal Android home app styled after Omarchy's waybar and walker: status bar, next event, notifications, recent apps, and a fuzzy app search where Jev picks the app by meaning when nothing matches by name. Project guide.
• Sippy (macOS water-break companion with optional Jev sip confirmation) - Open source · Free source build · BYOK. Small native macOS water reminder with a floating droplet buddy and menu-bar progress; Apple Vision face and hand tracking spots a possible sip on device, and with your own TypeSafe key Jev confirms the candidate frame before a drink is logged. Project guide.
• SiteClarity - Open source · Free · BYOK. Evidence-backed AI answer-readiness page audit with TypeSafe Jev quote-backed findings (hosted demo + MIT source). Project guide.
• skil.li - Open source · Free · BYOK. Native macOS launcher and assistant (apps, files, clipboard, snippets, windows, Hyper Key) that can also act on requests; its Ask router uses a fast decision model such as TypeSafe Jev to route what you typed before a larger model runs. Project guide.
• Skipto - Open source · Free source build · BYOK. Chrome extension that answers a question about a YouTube video by asking TypeSafe Jev a Choice over ~30-second transcript windows plus a yes/no “is it covered at all?” check, then highlights the progress bar and jumps to the best moment. Project guide.
• Slator Localization Request Readiness Inbox - Open source · Free · BYOK. Local inbox dashboard for enterprise translation intake: it reads unread Gmail or Outlook messages (read-only OAuth), uses Jev to identify translation requests, analyses supported source attachments, infers target languages and deadlines, and shows brief completeness, capacity and clarification points beside each email. Project guide.
• Slop Mop - Open source · Free · BYOK. Chrome extension that judges LinkedIn post writing with TypeSafe Jev (not an AI detector); free hosted use with daily check limits, or self-host MIT with your own key. Try app · Project guide.
• SlopFilter (Chrome extension that folds or hides AI-generated slop on X, judged by Jev and local rules) - Open source · Free source build · BYOK. Chrome extension for X: local rules scan each post, then Jev answers three questions (is it AI slop, what kind, does it show a human touch) and posts above your threshold are folded to one line or hidden. A floating counter shows how many were caught, you can pick lenient, balanced or strict, and marking a false positive restores the post and stops filtering that author. Without a key it runs on local rules only. Interface in Chinese. Project guide.
• Smart Dark (adaptive dark mode with optional Jev screenshot checks) - Open source · Free source build. Chrome extension that darkens websites by using a site's own dark theme when it has one and otherwise rewriting its colours while keeping layout, hover states and brand colours; optional Jev screenshot checks decide whether a page should stay dark or return to its original look. Project guide.
• Smart Find - Open source · Free source build · BYOK. Ctrl+F for questions: TypeSafe Jev scores page passages and highlights the answers (BYOK). Project guide.
• Smart Paste - Open source · Free source build · BYOK. Experimental Chrome extension that uses Jev to select and verify exact source text for web form fields, with paste and undo. Project guide.
• SmartDoc Vision (self-hosted archive and drawing review platform with Jev classifiers) - Open source · Free source build · BYOK. Chinese-language self-hosted review platform for archives, drawings, scans and text: a Go and Gin backend with MySQL and Redis and a Vue 3 admin console run configurable, traceable review pipelines on general LLMs and on Jev Choice and Noul classifiers, with model failover, circuit breaking and token accounting. Project guide.
• Snake × Jev - Source available · Free source build · BYOK. Browser Snake played by TypeSafe Jev with a live panel of every call — the state sent, the one choice question asked and Jev's probabilities; code only describes what the snake sees and where the food is, shuffles option order each turn and never overrides Jev's move. Distinct from the listed jev plays snake. Project guide.
• SoluTerminal (Windows workspace for Claude Code and Codex, with Jev smart approvals) - Open source · Free. Windows workspace that runs Claude Code and Codex side by side in PowerShell and WSL panes, keeps them running when the window closes, lets one agent ask another for a second opinion, and stops agents approving their own requests; Jev classifies each task into a tier and answers cautious permission prompts as smart approvals. Project guide.
• Soter - Open source · Free source build · BYOK. Discord moderation bot: TypeSafe Jev (OpenRouter) scores hate/spam; code deletes, flags, and can timeout. Project guide.
• SpaceSniffer, Jev edition (3D disk-space map for Windows that asks Jev which folders are safe to clear) - Open source · Free · BYOK. Windows disk-space map that draws your drive as a 3D city: every folder is a district sized by the space it takes. As soon as a folder is scanned, Jev labels what kind of folder it is, whether it only holds regenerable data such as caches, temp files and old installers, and how confident it is; safe-looking folders get a neon outline and land in a Reclaimable list. Jev never deletes anything; you confirm every deletion. Not affiliated with the original SpaceSniffer. Project guide.
• SplitSeek - Open source · Free. macOS desktop app that searches a folder of grey, flat LOG video footage by describing a moment in plain words: it samples a frame every 2 seconds, asks Jev a yes/no question about each frame, and ranks matching moments with timestamps, confidence and a preview from the best frame. Project guide.
• Sponsor Skip - Source unverified · Pricing unverified · BYOK. Locates YouTube sponsor reads with Jev-selected transcript boundaries and optional Deepgram audio analysis, with playback skipping. Project guide.
• SpotAsk - Open source · Free. Native macOS menu-bar assistant: summon it with a hotkey to ask your own model (BYOK), act on selected text in any app, or route a question to web AI, desktop apps or local CLI agents, with an optional Jev decision that picks the destination and shows its confidence before sending. Project guide.
• Stack4That - Open source · Free source build · BYOK. Describe what you are building; TypeSafe Jev assembles a coherent technology stack architecture. Project guide.
• Stillwatch - Open source · Free source build · BYOK. Notification guardian where TypeSafe Jev (via NanoGPT) picks out what needs you now from all your apps. Project guide.
• stream95 (Windows 95-style live-stream overlay whose chat is moderated by rules and Jev before it shows) - Open source · Free source build. Live-stream overlay for old-school programming channels built with React95: eight ready OBS scenes, chat that passes local rules and then Jev before it appears on screen, an LLM fallback when Jev is unsure, a Clippy that answers chat questions and optional text-to-speech. Portuguese-language docs. Project guide.
• Surge Desk - Source available · Free · BYOK. Hackathon demo for 911 surges: each call transcript (multilingual, with location) gets six typed TypeSafe Jev questions in one request — same emergency as an open incident?, what the call adds, severity, life threat — so code can merge duplicates, keep new life-safety facts, split hidden emergencies and rank P1–P4, asking the supervisor when Jev is under 70% sure; nothing is dispatched without human approval. Project guide.
• Switchyard - Open source · Free source build · BYOK. macOS menu-bar Dia/Chrome profile router: TypeSafe Jev chooses the track, then learns local rules. Project guide.
• Synax - Open source · Free · BYOK. Open-source local workspace for coding agents and codebase docs (agent chats, files, diffs, terminal) whose Electron app offers optional Jev-assisted Computer Use: with a TypeSafe key (or OpenRouter), Jev picks bounded semantic actions for the app-hosted Cua Driver instead of the default Direct Cua strategy. Project guide.
• Tab (Mac next-action autocomplete) - Source available · Free source build · BYOK. Hackathon proof of concept that predicts your next action on a Mac and shows it as a gray ghost (a highlighted button or gray text); press Tab to do it. Jev picks among code-proposed candidate actions, and x402 paywalls become testnet payments with an amount check. Project guide.
• Tab Bouncer - Open source · Free source build · BYOK. Chrome extension that rates open tabs for a typed task with TypeSafe Jev (keep noul + kind choice; batches of 120), then closes mismatches. Distinct from Jev for Chrome. Project guide.
• Tab Land - Open source · Free source build · BYOK. Chrome extension that finds any tab fast, groups tabs by the job you are doing, and closes stale ones; with a Jev key, Jev double-checks each group, removes tabs that do not fit, and names groups. Project guide.
• TabTidy - Open source · Free source build · BYOK. Chrome MV3 extension that filters tabs locally (active, pinned, audible, downloading stay), then sends one Jev request with a KEEP / CLOSE / UNCERTAIN choice question per remaining tab; it closes only what you confirm (one-click undo through Chrome sessions) and turns your corrections into domain rules that override Jev next time. Project guide.
• Tabwerk - Open source · Free. German-language Chrome extension for people with many tabs: quick search, groups, sorting, duplicates, page watchers and full history, where exact work is done in code and optional Jev calls answer the semantic questions (find a tab by description, suggest groups, rank priority, judge watcher conditions). Project guide.
• talktojev (Jev Prime) - Open source · Free · BYOK. Generation-as-classification web chat: TypeSafe Jev chooses every word via OpenRouter Decisions (no LLM). Try app · Project guide.
• Tally (automatic time, income and evidence tracking for hourly contract work) - Open source · Free source build. Menu bar app for hourly contractors that records which app, window and tab you are in every few seconds, assigns time to projects by your rules, seals intervals into a hash-chained ledger and bills only explicit client sessions; optionally Jev assigns unmatched windows to projects (at 80% confidence or more) and to activity categories. Project guide.
• Tavli - Open source · Free source build · BYOK. Greek backgammon (Portes) played against TypeSafe Jev via OpenRouter, with a typed-decision tutorial. Project guide.
• Tessmora (MMA-RAG) - Source available · Free source build · BYOK. Tessmora (Chinese docs): self-hosted omni-modal agentic retrieval platform for documents, images, audio and video, with opt-in TypeSafe Jev semantic judgment for a simple-question intent fast path, choice reranking and citation-support diagnostics. Project guide.
• Text-Only YouTube (userscript with Jev genre tags) - Open source · Free. Violentmonkey userscript that turns YouTube into a plain list of titles without thumbnails, avatars or Shorts; one click sends the visible titles to Jev to tag each video with a genre icon you can filter by. Project guide.
• Text-To-VRMA - Open source · Free · BYOK. Text-to-motion VRM avatar app (VRMA/GIF/video export) where optional TypeSafe Jev analysis of emotion, gaze and tempo conditions ARDY/LLM motion generation. Project guide.
• tg-crush - Open source · Free source build · BYOK. Local real-time Telegram coach: TypeSafe Jev scores drafts and partner messages before you send (BYOK). Project guide.
• TG-Spam - Open source · Free source build · BYOK. Self-hosted Telegram anti-spam bot and Go library (maintained since 2023); its optional Jev provider asks one typed spam question (plus an optional gibberish question) and thresholds the returned probability. Project guide.
• THE SWARM - Open source · Free source build · BYOK. WebGL swarm of 1,000 light particles whose strategy (flee, attack, protect, split, regroup) is chosen live by TypeSafe Jev every 1.5 seconds or on events, while local rules move each particle. Project guide.
• Tidyward (private file housekeeper for macOS where Jev answers what each file is) - Open source · Free source build · BYOK. Early macOS command-line housekeeper: it identifies files locally first, masks real names before anything leaves the machine, asks Jev only semantic yes/no questions such as "is this an invoice?", and your plain-text rules turn the answers into a keep, archive or Trash plan you confirm; every move can be undone. Project guide.
• TipTour - Open source · Free source build · BYOK. macOS menu bar app that uses Jev to select desktop click targets from typed requests, alongside a separate Gemini voice mode. Project guide.
• Toru (recording infrastructure for coding agents, with optional Jev browser navigation) - Open source · Commercial · Freemium · BYOK. MCP recorder that lets Claude Code, Cursor, Codex and other agents record product demos, changelog GIFs and bug repros in your signed-in Chrome or on your Mac, renders them with your brand and checks each take; optional server-side Jev navigation picks browser operations and targets for a goal. Project guide.
• ToS Decipher (terms-of-service and privacy-policy reviewer whose claims are checked by Jev) - Open source · Free source build · BYOK. English and Spanish web app, Chrome/Edge MV3 extension and Python API that summarise terms of service and privacy policies, highlight important clauses next to their source text and find policy links on a page; Jev classifies links and verifies each finding against the source passages. Project guide.
• ToS Watch - Open source · Free source build · BYOK. Email alerts on material ToS/privacy changes via Open Terms Archive + TypeSafe Jev (tos.watch). Project guide.
• Transcript Lens - Open source · Free source build · BYOK. Next.js (Türkçe UI) YouTube transcript explorer: TypeSafe Jev classifies blocks for kind/value/signals without rewriting text. Project guide.
• Treg ("OpenRouter for agent tools": pay-per-call tool catalog whose tool search is judged by Jev) - Source available · Paid · Commercial. Hosted tool registry from Superdesign: agents use one base URL and token to call thousands of priced endpoints (SEO, social, people and company enrichment, ads, scraping, media generation) without provider accounts, plus team keys and skills; tool search sends the query and candidates to Jev in one judge request, and Jev is also offered as a catalog endpoint. Project guide.
• Tube Focus Mode - Open source · Free source build · BYOK. Chrome extension that filters YouTube down to your channels, playlists and chosen topics: Jev scores each other video against your focus topics, and off-topic videos are blurred or blocked. Project guide.
• TypeSafe Fun AdBlocker - Open source · Free source build · BYOK. Experimental Chrome extension that asks Jev whether heuristically selected DOM elements are ads, then removes or highlights matches. Project guide.
• uCable (your YouTube subscriptions as an early-2000s satellite TV guide, with Jev tagging every video) - Open source · Free source build · BYOK. Web app that takes the last 14 days of videos from your YouTube subscriptions and schedules them into numbered themed channels with a two-screen TV guide; type 1-0-5 to change channel. Jev labels each video with its channel theme, while a plain-code scheduler decides order and remembers where you stopped. Italian UI. Project guide.
• Unclutter - Open source · Free source build · BYOK. Classifies page clutter with Jev and saves reversible hiding rules for similar pages, with manual analysis by default. Project guide.
• undertone - Source available · Free source build · BYOK. Text box that reads draft tone with TypeSafe Jev before you send (hosted demo + source; no LICENSE file). Project guide.
• Usher (macOS menu bar app that sizes the focused window by the kind of work, classified by Jev) - Open source · Free source build · BYOK. When you switch windows, Usher sends the app name and window title to Jev on OpenRouter, sorts the window into Code, Reference, Media, Writing, Chat or Ambient, gives it that category's share of the screen (two-thirds for code, full for media) and arranges the other windows around it. Shares are adjustable from the menu bar, answers are cached per app and title, and with no network or key it guesses from the app. Project guide.
• UtilityPot - Open source · Free. Free web utilities for messy data, shared text, diagrams and planning: browser-local CSV/JSON conversion, date cleanup and meeting-time finder, plus three bounded Jev tools for custom entity extraction, text-to-Mermaid flowcharts and finding the exact paragraphs on a webpage that answer a question. Project guide.
• Varina - Source available · Free source build · BYOK. Chinese-language multi-seat AI design-exploration workbench (local web UI + CLI) where an optional TypeSafe Jev post-gate decides START or ASK — whether a finished answer merits deeper multi-seat exploration — before any extra model spend; Jev never writes answers. Project guide.
• VerdictKit - Closed source · Freemium · Commercial. Independent hosted Jev playground: build Choice, Score, or Noul questions over your text or JSON and inspect typed responses. Project guide.
• Vertika Command - Closed source · Free · BYOK. Chrome extension that runs your browser by voice or typed commands; free in-browser Laya or TypeSafe Jev online with your own key. Project guide.
• Vibe Check for X - Source unverified · Pricing unverified · BYOK. Chrome extension that uses Jev to score draft X posts and display a verdict, with optional OpenAI media descriptions. Project guide.
• Vicaura - Closed source · Pricing unverified. Hosted product-to-markdown tool: extract features, ICP, messaging, and pricing strategy into md for coding agents (vendor: powered by Jev). Project guide.
• voice-browser - Open source · Free source build · BYOK. macOS command bar that drives Chromium/Dia/Arc/Chrome from plain English with TypeSafe Jev (or SubQ Snap) picking clicks and targets. Project guide.
• Vonage Video API x Jev support demo - Open source · Free source build · BYOK. Vonage Community demo where Jev reads a customer's chat messages for sentiment and frustration and escalates to a live Vonage video call with a human agent. Project guide.
• Warren Duffer - Open source · Free source build · BYOK. Intraday Nifty-50 trading bot where TypeSafe Jev ranks names and code sizes/stops live broker orders (experimental; real-money risk). Project guide.
• Watermelon - Open source · Free source build · BYOK. Status-update honesty auditor: TypeSafe Jev judges language while code parses slip signals; live demo available. Try app · Project guide.
• X Semantic Filter (Chrome extension that folds low-value X posts by your plain-language rules, with Jev) - Open source · Free source build · BYOK. Chrome extension that folds posts on X in real time: describe what you don't want in plain language, every post gets a probability per rule, and anything above your threshold collapses into a one-line bar you can expand. Backends are Jev, Cloudflare's Clef-omni (which also reads images) or an OpenAI-compatible model; bilingual Chinese and English UI, no backend of its own. Project guide.
• Xift - Open source · Free source build · BYOK. Firefox add-on that collapses noise on your X timeline: TypeSafe Jev (via OpenRouter) scores each post for engagement bait, rage bait, low substance, and self-promotion, and posts over your thresholds fold to a one-line reason. Project guide.
• XPostFilterPlus (Chrome extension that hides X posts by keyword or by Jev classification) - Open source · Free source build. Unofficial Chrome extension that hides X posts using keyword rules with case, accent and symbol normalisation, plus optional semantic exclusion conditions judged by Jev, OpenAI Decisions or Cloudflare Clef with your own key; settings open over the X page and a chip counts hidden posts; Japanese and English documentation. Project guide.
• Xtags - Open source · Free source build · BYOK. Chrome extension/userscript that tags X posts with TypeSafe Jev intent and risk signals for personal local browsing. Project guide.
• YapSketch - Open source · Free source build · BYOK. macOS app for explaining bugs by talking and pointing: record your screen and voice, and Jev splits what you said into points, matches each to what you showed, and picks the clearest screenshot, then puts numbered text plus images on your clipboard for a coding agent. Project guide.
• YouTube Jev Filter (Chrome extension that hides clickbait, hype and Shorts on YouTube using Jev) - Open source · Free source build · BYOK. Manifest V3 Chrome extension: for each video card it sends the title, channel, length and description to Jev with three questions (clickbait yes/no, hype yes/no, substance score 0-3) and hides or dims the video when a threshold is crossed, with soft, normal and strict modes. Shorts are removed. Answers are cached per video ID; the README estimates about 3 cents per 1,000 videos. README in Russian. Project guide.
• YouTube Sponsor Skip (Jev) - Open source · Free · BYOK. Chrome/Safari extension that finds spoken sponsor segments in Russian YouTube videos via TypeSafe Jev on subtitles and skips them. Project guide.
• Zearch - Open source · Free source build · BYOK. Perplexity-style open-source UI for TypeSafe Jev: search to evidence state to typed decisions. Project guide.
• Zéro Ping (local inbox for lead developers ranked with Jev) - Open source · Free source build. Local inbox that replaces GitHub notification emails with a ranked list of reviews, replies, PRs and sprint items; Jev ranks lists and gives each PR KPI cards such as production risk, review effort and merge readiness. Project guide. Applications with a user-facing workflow powered in part or entirely by Jev. The full app directory shows platforms, access requirements, and Jev's specific role. Tags distinguish source access from pricing; commercial and closed-source apps can qualify. Makers are welcome to submit their own apps.

Developer projects and integrations

• 2048 with Jev - 2048 experiment where a bot chooses moves via the TypeSafe SDK (typesafe-sdk, pinned) using a fixed Jev model id so runs do not silently follow jev-latest alias moves. Project guide.
• 2048: Jev vs Laya - Experiment where two System One decision models — hosted TypeSafe Jev and open-weight Laya (plus a LoRA fine-tune) — play 2048 on the same seeds against random and greedy baselines, each move asking a move choice and a danger yes/no from a board description, with every move logged and reports comparing win likelihood, choices and latency. Project guide.
• 8-K Filings and Decision Models - Jev rated 26,962 S&P 500 8-K filings (2013 to 2026) from very bad to very good for $2.32; within item group and year the rating's rank correlation with the next two days' market-adjusted return is +0.025, positive in 13 of 14 years, and the move is mostly gone by the next open. Project guide.
• @amatelic/jev - Node library and local server for one-line choice/noul/score decisions from small open models, with no API key or cloud call. Project guide.
• @contexera DSH jev plugin - DeepSeek Harness plugin monorepo whose jev package adds a TypeSafe Jev service for typed decisions alongside channel-gateway and context-continuity plugins. Project guide.
• @rbxts/jev (Jev for roblox-ts) - Typed TypeSafe Jev client for roblox-ts games: declare choice, score and noul questions about one state, and the option names become TypeScript types your server code branches on, with a shared rate limiter, retries and request deduplication over Roblox HttpService. Project guide.
• @sideform/browser-agent - TypeScript goal-driven browser agent library whose step policy uses TypeSafe Jev via a System One client. Project guide.
• A Decision Model as a World Model - Code, data, run logs and manuscript for a KIIS 2026 Fall Conference paper testing whether a frozen TypeSafe Jev decision model can serve as an LLM agent's world model in TextWorld, against zero-shot and LoRA-tuned Qwen3-4B in decision-style and generative set-ups; the author reports Jev near the oracle on task success without task training. English and Korean docs. Project guide.
• Abide - Enforce AGENTS.md/CLAUDE.md rules on every coding-agent edit/turn with TypeSafe Jev (or Vercel AI Gateway) per-rule probabilities on diffs. Project guide.
• Abstract Sift (screen research abstracts against your interests with Jev, OpenAI Decisions or both) - Local CLI for screening conference or paper abstracts against a research-interest text with Jev, OpenAI Decisions or both side by side, with agreement and disagreement reports. Concurrent, resumable scoring, cost estimates and a --max-cost-usd cap, CSV/JSONL exports and an offline HTML report; an ESMO 2026 run guide is included. Project guide.
• Accurate and Useless (article and anonymized data from 5,300 Jev calls in a multi-agent coding pipeline) - Article, figures and anonymized data from a 16-day trial of Jev inside a multi-agent build pipeline for Claude Code, testing advisory checks, an on-demand tool for agents and binding gates. Jev answered 4,222 calls across five real builds plus 1,092 offline; the authors report that most label sets had a single class, so no check showed reliable judging and the hoped-for token savings were never measured. recompute.mjs rebuilds all 103 published numbers from the data. Project guide.
• ACP Bridge - Bridge that exposes local coding agents over the Agent Client Protocol, with a virtual auto agent where TypeSafe Jev picks the best-fit agent per task with a confidence gate. Project guide.
• adecider - Typed System One decisions for coding agents: one call, many Noul/Choice/Score questions, pluggable local Laya / TypeSafe Jev / OpenAI-compatible backends. Project guide.
• adk-go-typesafe - TypeSafe System One Go client and Google ADK-Go function tool (OpenAPI-generated types; Apache-2.0). Project guide.
• Advocaat - TypeScript ask client that batches typed Jev choice, score, and yes/no questions about structured data, with optional Vercel AI Gateway support. Project guide.
• Agastya (training-free retrieval where Jev walks a document's table of contents and local embeddings rank the leaves) - Retrieval research code: Jev walks a document outline level by level, giving one noul probability per candidate section, and EmbeddingGemma ranks the leaves it keeps. No training or labels; the README reports the maintainer's benchmark against flat Jev ranking and embedding baselines. Project guide.
• agemem (project memory for coding agents, with Jev screening notes) - Experimental memory for coding projects: a readable Markdown memory bank, screened SQLite notes and cited source excerpts, exposed to Codex, Cursor or Claude Code over MCP; decision models screen, classify and select memories, with Jev as the default and local QSIV or ollaya models as options. Project guide.
• Agent Browser (caglarkc) - Research browser runtime for AI agents: observes pages through Lightpanda DOM controls, lets Jev choose the next action, and verifies outcomes independently, with reproducible benchmarks. Project guide.
• Agent checks (routing, brief lint, plan and review gates) - Claude Code hooks and scripts that put Jev judgements around AI coding agents: route each turn on purpose, lint briefs before dispatch, block inline edits to files assigned to a worker, check cited sources in research reports, and ask whether returned work shows a check actually ran. Project guide.
• Agent Chess (Claude vs Jev) - Local chess board styled after chess.com where Claude models and Jev play each other, with per-side engine settings, Jev move probabilities in the live view, a per-run spend cap and server-side legality checks. Project guide.
• Agent Router - Quota-aware Herdr launcher that uses TypeSafe System One (Jev) to pick Cursor/Claude Code/Codex/OpenCode model and effort after local eligibility rules. Project guide.
• Agent Seek - Cheap web recall for agents: You.com discover plus TypeSafe Jev cascade ranking (REST/MCP/UI); does not write answers. Project guide.
• Agent Stack - Local multi-agent software team stack where TypeSafe Jev makes small typed merge/assignment decisions (no root LICENSE at tip). Project guide.
• agent-chaperone - Calibrated firewall for agent tool calls: MCP proxy plus hooks adapter; TypeSafe Jev screening with shadow mode and a local judgment log. Project guide.
• agent-desktop - Rust macOS accessibility CLI for desktop computer use; optional jev-desktop skill/scripts use TypeSafe Jev for target/command choice without putting the a11y tree in agent context (BYOK). Project guide.
• agent-doc-search - Docs routing for coding agents: BM25 and a TypeSafe Jev search propose, a Jev yes/no check confirms relevance. Project guide.
• agent-evals - Deterministic LLM-agent eval harness with optional calibrated TypeSafe Jev judge for CI deploy gates. Project guide.
• agent-factory (self-hosted company of AI agent seats that review and land code under scripted gates, with Jev routing) - Bun/TypeScript service with PostgreSQL that keeps your git repositories moving: agent seats review pull requests, triage issues, dispatch fixed jobs to disposable runner boxes and publish bounded results back to the forge, under hard scripted gates and a daily token budget kill switch. Roles name a capability tier rather than a model, and Jev picks the route. Project guide.
• agent-fastpath - MCP decision layer for coding agents: deterministic rules then TypeSafe Jev ship/risk/triage/browser tools. Project guide.
• agent-on-agent-ids - Intrusion detection for AI agents over Opik traces, with TypeSafe Jev as a supported low-cost judge. Project guide.
• agent-reflex - Coding-agent plugin that uses a System One decision model (Jev, or a local model served by llama.cpp, Ollama or LM Studio) for a shell guardrail, post-edit checks, task complexity classification and a confidence check across OpenCode, Claude Code, Codex, Hermes, Qwen Code, Antigravity and Pi. Project guide.
• agent-sdk-2048 (two Agent SDK players race 2048 in an emulator: a normal model turn vs Jev decide()) - Side-by-side race in mGBA: one Agent SDK player takes a normal model turn per move, the other uses Jev decide() to make a typed choice among up, down, left and right, with traces of what Jev saw. Without a key the Jev player refuses to run unless you pass a labeled stand-in. Project guide.
• agent-test-pruning - Scripts and data for the paper "Which Tests to Keep When a Coding Agent Writes the Test Suite": replays of Claude Code writing tests, planted and real bugs, and pruning signals including four facet questions asked of Jev per test. Project guide.
• agent-tree - Claude Code plugin that only watches: optional bottom HUD lines merged into your status line plus an on-demand /agent-tree pane showing the main agent and subagents, and — when you opt in — TypeSafe Jev reading a small session summary to draw four advisory flags (looping, stuck, fan-out wide, near done) and a health score. Project guide.
• Agentic Failure Attribution (what agent traces do and don't tell you, with Jev step judgments) - Study on τ-bench banking (20 models, 7,760 conversations, 97 tasks) testing the signals trace-QA tools rely on against real outcomes; Jev judged each step and every user message (17,160 calls), and the finding is that task difficulty predicts failure better than anything in the trace, retrieval matters within a task, and user pushback is not an early warning. Includes a trace dashboard. Project guide.
• agentic-bees (agent-directed Btrfs deduplication CLI with an advisory Jev next-step check) - Rust CLI and MCP interface that lets an AI agent inspect Btrfs storage, run bounded scans, review matches and apply kernel deduplication inside an authorized scope. An optional decide command asks Jev for an advisory next step using only allowlisted counters. Project guide.
• AgentRun - Workflow language for the agents you already run: define repeatable steps, let TypeSafe Jev make focused typed checks (@parcha/agentrun-jev), and call an agent only when work needs investigation. Project guide.
• AgentScope (multi-agent framework with a built-in Jev classifier model) - Popular Apache-2.0 Python framework for building multi-agent applications. Since 2026-09 it ships a classifier interface with JevClassifierModel, which turns AgentScope questions into TypeSafe typed questions through the TypeSafe SDK, so agents can hand narrow judgments to Jev instead of a chat model. Project guide.
• AGIW MCP (careful coding workflow on local models, with consent-gated Jev checks) - Local MCP server for Claude Code, Factory or OpenCode: a readiness check, a Nisi workflow where one local model drafts, tests run and a second reviews, and optional advisory Jev questions before and after a task (risk, missing detail, unbacked claims) that send nothing until you consent. Project guide.
• AGNES Arena (Pokémon battles with Jev-assisted decisions) - Pokémon Showdown battle arena where two persistent agent players evaluate legal moves through Jev, commit an action and play whole matches, with session management, a scheduler and a browser battle viewer. Project guide.
• agrouter - Go CLI that asks Jev which coding-agent CLI, model and reasoning effort (Claude Code, Codex or any configured CLI) a prompt needs, then prints the choice as JSON or launches that CLI with it. Project guide.
• agy-cua (background computer use for Antigravity) - Unofficial MCP plugin that gives Google Antigravity (app and agy CLI) background computer use on macOS through the open-source Cua Driver, with an optional Jev lane that picks each guarded step in about 0.2 to 0.3 s and refuses secure fields and send or commit actions. Project guide.
• agy-jevgate - Fail-closed Antigravity PreToolUse hook for shell commands using TypeSafe Jev risk scores. Project guide.
• AI PM Job Radar - Every open PM role at 36 AI and tech companies, sorted by Jev (via Vercel AI Gateway) into AI PM without ML, AI PM with ML, non-AI PM and not-a-PM-role. Project guide.
• AI-Core Engineering Workbench (15-lesson course code ending in a guarded issue-to-patch workflow with Jev routing) - Runnable companion to a 15-part bilingual course on anirach.com: lessons go from plain functions to a guarded issue-to-patch workbench where Jev gives a typed route and bypass judgment, checks inspect the candidate patch, and a separate quality gate reviews the explanation. Runs offline with scripted fixtures by default; real Jev calls are optional. Project guide.
• aidiff-gate - Claude Code plugin: optional TypeSafe Jev judge flags flaky AI-written tests and trims walls of text. Project guide.
• Ailuros - Ingress decision service for proxies: Envoy (ext_proc, ext_authz), NGINX auth_request, Traefik ForwardAuth and Envoy Gateway on Kubernetes send each request to Ailuros, which asks TypeSafe Jev or an OpenAI-compatible model, including local models, for an allow or block decision. Project guide.
• airportiq (airport capacity-pressure agent with a Jev guardrail and Haiku fallback) - One-day take-home: a conversational agent over public FAA and BTS data that shows which US airports are under the most capacity pressure. Numbers come from deterministic code; an LLM calls typed tools, and a guardrail asks Jev first, falling back to Claude Haiku when Jev is unavailable or unsure. Project guide.
• aiscript - CLI that turns a plain-English intent into a cached Bash or PowerShell script written by an LLM, then checks every generated script with a static scan and optionally Jev (or a local Ollaya model) before it is saved or run; risky scripts need an explicit approval. Project guide.
• aisdk/typesafe (native Jev decisions provider for the framework-agnostic PHP AI SDK) - Provider package for the PHP AI SDK (aisdk/core) that runs Jev decisions natively from PHP 8.3+: portable predicates map to TypeSafe noul, choices and scores to native criteria, and values, probabilities, confidence, usage and the raw response are kept. Works with any PSR-18 HTTP client. Project guide.
• aitomaton (error-to-fix automation with a Jev decision tier) - Service that listens to Sentry, urgentry and Alertmanager, analyses each error alone, in bursts and as long-term patterns across a Docker Swarm, and decides what to do through a chain of rules, the open Laya model, Jev and an LLM; code defects can go to a coding agent that opens a pull request. Project guide.
• AIWG - Multi-agent workflow framework for existing AI tools whose opt-in decision runtime can bind the same validated decision definitions to Jev or to an LLM subagent, with pins, projection, bounded retries, fallback and receipts. Project guide.
• ajevt-browser - Bounded TypeSafe Jev browser tool for Pi, OpenCode V2, Amp, and MCP powered by agent-browser (AGPL-3.0). Project guide.
• Albanian sentiment analysis with Jev - Reproducible study classifying Albanian COVID-19 Facebook comments (AlbAna, three-annotator labels) with TypeSafe Jev and no training, re-running the 20 classifiers of Kastrati et al. (2021) on the paper's own test split; maintainer reports Jev at 75.71 weighted F1 vs 72.08 for mBERT, with paired-bootstrap intervals. Project guide.
• AlphaOptimizer - Compact large Codex/tool outputs locally; optional TypeSafe Jev ranks relevant chunks (BYOK). Project guide.
• Amplifier Jev provider (Microsoft) - Optional, experimental provider module for Microsoft's Amplifier agent framework that exposes Jev decisions (predicate, choice and score) to consumers such as browser and computer UI use, without ever joining the chat fallback chain. Project guide.
• AnchorLint - Audits internal links in built HTML sites with deterministic checks plus optional TypeSafe Jev promise/relevance judgments. Project guide.
• android-jev (FZ2000) - MCP server/skill: drive an Android phone over adb with TypeSafe Jev choosing each next action (distinct from listed jev-android / jevdevice). Project guide.
• Ankan Core (self-hosted anything tracker with an MCP proxy and optional Jev routing for Telegram messages) - Self-hosted tracker for anything you want to log, with a Vue frontend, a local Convex backend and an MCP proxy, started with Docker Compose. A Telegram bot can add entries from chat; with a Jev key set, incoming Telegram text is classified by Jev before falling back to Google AI. Email one-time-code sign-in through ZeptoMail. Project guide.
• anofox-decide - DuckDB extension: NL predicates and answer sets via TypeSafe Jev or local open System One–compatible models (remote opt-in). Project guide.
• Antigravity System One (agy-smart: Jev picks which skills the Google Antigravity CLI loads per prompt) - Wrapper for the Google Antigravity CLI (agy) that classifies each task with Jev and mounts only the relevant skills before handing off to the main reasoning model, instead of loading 50+ installed skills (3,000-8,000 tokens) into every turn. Includes a local dashboard. Project guide.
• antigravity-approval (fail-closed tool-call guardrail for Antigravity: local rules first, Jev for the ambiguous calls) - PreToolUse permission hook for Antigravity: a fast local tier (whitelist, blocklist, network gate, path guard) settles clear commands with no network traffic, ambiguous ones go to Jev, and anything uncertain falls back to asking you. Also hosts the shared jev_eval policy core reused by the author's ZCode hook. English and Chinese docs. Project guide.
• AnyJev (Nokia Applied Research) - Turn open LLMs into Jev-style typed decisions with probabilities (L0–L2); independent of official Jev / TypeSafe hosted API. Project guide.
• ApexBench (benchmark for forecasting Apex Legends team placements from replay data, Jev by default) - Open benchmark that loads a historical ALGS Apex Legends replay, rebuilds the match state at a chosen time and asks a model to forecast one team's final placement, then compares forecasts with the result. Jev is the default model through OpenRouter's Decisions API; GPT, Claude, Gemini, DeepSeek and other OpenRouter models can be swapped in. Includes terminal playback, caching and export. Project guide.
• App Store Indexor (describe an app in plain words and get matching US top-chart apps scored by Jev) - Natural-language lookup over Apple's public App Store charts: describe an app or its look ("habit tracker with streaks", "yellow bird icon"), local text and icon embeddings shortlist up to 120 apps, and one Jev request scores them; apps under 30% are dropped. Every field is marked verified, estimated or unavailable. Runs as a local Next.js site. Project guide.
• Apple RAG MCP - MCP server for Apple developer documentation and WWDC transcripts: hybrid retrieval, then TypeSafe Jev ranks the merged candidates against the query's API, platform and version needs, falling back to a Qwen3 reranker; hosted with a free tier and paid Pro plan, or self-host the MIT source. Project guide.
• argmax - Unity swarm-programming RTS where four commanders, including TypeSafe Jev answering typed battlefield questions, emit the same decision type. Project guide.
• article-quality-gate (build-blocking quality gate for articles with deterministic checks and a Jev judge) - Quality gate for a site's articles: deterministic structural checks plus a 12-question judge, a hash ledger that forces re-scoring after edits, calibration against golden samples, and a build step that blocks publishing below the bar. The judge can be a mock or Jev. Project guide.
• arXiv Jev Digest (daily arXiv screening with Jev, summaries by email) - GitHub Actions job that has Jev judge each new arXiv paper in your field against your research interests, filters and ranks by relevance and verified author institutions, then has DeepSeek write structured summaries and emails the top picks. Project guide.
• arxiv-relevance-watch - Zero-dependency Node CLI that screens arXiv for one research topic and ranks papers with five typed questions to a System One decision model (Jev via OpenRouter, TypeSafe or a local /v1/systemone), with no generative model anywhere and a second stage for reading survivors against your open items. Project guide.
• ask-jev (Aether-254) - MCP + Codex/Claude plugin for TypeSafe Jev evaluate/batch/ping (Choice/Score/Noul). Project guide.
• ask-jev (agent skill and zero-dependency runner for typed Jev decisions in Claude Code, Codex and other agents) - Agent skill that tells a coding agent when and how to ask Jev for typed selections, labels, yes/no checks and scores instead of prompting and parsing text, with a Node.js runner that has no package dependencies. It can rank files and check claims, and works with TypeSafe, Jev on Cloudflare Workers AI or the Vercel AI Gateway. Includes a trigger-evaluation harness with saved results. Project guide.
• ask-jev (Go CLI and MCP server) - Standard-library Go CLI and MCP stdio server for TypeSafe's System One API: ask choice, yes/no (noul) and score questions or a batch, get JSON back, and keep a local outcome log. Project guide.
• ask-jev-skill - Hermes skill that asks TypeSafe Jev for typed Choice/Score/Noul tiebreaks when paths remain plausible. Project guide.
• ask-jev-skill (DiegoSalazar) - Claude Code skill plus a dependency-free Bun jev CLI (yes, pick, rate, rank, triage) that lets Claude offload cheap, repetitive judgments to TypeSafe Jev — which files to read, whether command output means pass or fail, which skill fits, filtering and ranking — so raw content stays out of Claude's context. Project guide.
• AskGLP1 (paper code: GLP-1 patient-narrative knowledge graph with Jev post screening) - Code, prompts and evaluation questions for a paper that turns Reddit, WebMD and FAERS data about GLP-1 medications into a source-linked knowledge graph and a KG-grounded question-answering agent; Jev is the screening classifier that keeps posts reporting a drug, a personal experience and clinical detail before GPT extraction. Project guide.
• askif - TypeScript toolkit for typed branching on Jev answers — ask.if(…).elseif(), ask.switch, ask.score and .unsure() bands — with a provider-neutral Backend contract, a mock backend for tests, and @askif/jev as the TypeSafe Jev bundle. Project guide.
• askjev - Unofficial MCP server for TypeSafe Jev (local stdio or hosted Worker): agents get calibrated Noul/Choice/Score answers over held context. Project guide.
• assay - Local open take on System One/Jev: typed answers with calibrated confidence from a small laptop model (stdlib Python). Project guide.
• Astra-Ares - Adaptive GPT-6 Astra reasoning-effort selection during Codex tasks via TypeSafe Jev (experimental patched Codex CLI). Project guide.
• astrbot_plugin_at_gate (AstrBot plugin that merges split messages and silences repeated @ pings, with Jev deciding) - AstrBot chat-bot plugin for two common group-chat problems: users who send one sentence in several pieces (it waits and merges them, images included) and users who keep @-ing the bot while a long answer is generating (it stops the queued "I'm here" replies). Whether a fragment is finished or a ping needs a reply can be decided by Jev, another configured model, or offline rules. Chinese-language docs. Project guide.
• astrbot_plugin_jev_gate (AstrBot plugin: Jev decides when a group-chat bot should speak up) - AstrBot plugin that has Jev score each group-chat message and triggers one proactive reply only when it is worth joining in, with persona, questions, thresholds, cooldowns, daily caps and a dry-run audit mode; the reply itself is still written by AstrBot's main LLM. Project guide.
• astrbot_plugin_jev_router - AstrBot plugin: TypeSafe Jev picks the sub-agent (model, persona, tools) for each chat message. Project guide.
• astrbot_plugin_smart_wake - Chinese AstrBot plugin that lets a group-chat bot join conversations without being @-mentioned: topic keywords trigger a weighted dice roll, then TypeSafe Jev judges reply value × naturalness × heat × energy before waking the bot; Jev also debounces when a conversation ends and confirms 'please be quiet' requests to enter a muted state, with the last 200 decisions in the dashboard. Project guide.
• Athena Research (pre-registered experiments on AutoML decisions, including whether Jev can choose explanation methods and search depth) - Evidence repo for articles on LLM-assisted decisions in tabular machine learning: designs, pre-registrations, prompts, code, run records and statistics. The JEV study asks whether Jev can make bounded AutoML choices; richer semantic descriptions lifted one decision from 22.6% to 58.1% on the original benchmark, but on six unseen datasets nothing it added was significant after correction. Other studies compare chain, tree and graph prompting and semantic context with local and hosted LLMs. Project guide.
• atoma - Agent coordination platform (hosted atoma.run or self-hosted, AGPL-3.0) that turns a request into checked work; TypeSafe Jev takes bounded decisions such as which agent or recipe takes a task and whether a plan is approved. Project guide.
• AudioJev-Inference (Python library and System One-style HTTP server for the open AudioJev audio decision model) - Inference library and FastAPI service for AudioJev, an open audio decision model described in an arXiv paper: give it an audio clip and natural-language questions and it returns candidate probabilities, yes/no judgments or ordinal scores, several questions per request. The server exposes a POST /v1/systemone endpoint in the Jev request shape, and a standard-library client is included. Runs locally on an NVIDIA GPU. Project guide.
• auth-audit-jev - Observation-only Python EventPlugin: TypeSafe Jev audits allowlisted OAuth/IAM authn/authz events and emits advisory alerts (never enforces). Project guide.
• Auto Mode for Paseo - Paseo plugin: TypeSafe Jev (default) or local Laya selects Codex model, effort, mode, and speed per turn. Project guide.
• Auto TLDR - Claude Code mod that follows a detailed answer with a self-contained TLDR request in the same conversation (handy with per-reply Read aloud); optional Jev filtering asks TypeSafe for the probability that a summary would materially help and submits only at 0.7 or higher, sending the completed answer only. Project guide.
• AutoJev - Desktop local model-router gateway for Codex/Claude Code/Hermes; optional OpenRouter Jev decision routing with local fallback (AGPL-3.0-only). Project guide.
• autoloop - Config-driven GitHub issue → triage → implement → review → PR pipeline for coding agents, with an experimental, off-by-default Jev shadow mode (typesafe/jev-1.13 via OpenRouter) that evaluates triage and auto-merge decision points alongside the pipeline and only logs. Project guide.
• automodel - Claude Code proxy where Jev (via OpenRouter) picks the model and effort for each prompt, subagent and workflow step. Project guide.
• AutoRouter (claude-autorouter) - Independent local gateway that routes each Claude Code request across Haiku, Sonnet and Opus after an evaluator rates it — a local Ollama /v1/systemone model by default, or hosted TypeSafe Jev with --evaluator jev — checking model compatibility and context capacity first. Project guide.
• Axeon JEV Router (OpenClaw) - OpenClaw plugin that asks Jev to pick the model for each interactive turn (Luna, Sol or Astra by default) with owner-controlled /autoroute on, off or shadow, keeping host locks, fallbacks and thinking settings untouched. Project guide.
• Axiomaton (tell your model what changed since its training cutoff; Jev resolves contradictions) - Go command-line tool from Canonical that keeps append-only logs of dated facts about languages, tools and libraries and renders those newer than a model's knowledge cutoff as Markdown for its context; logs for Go, Python, Rust, Zig and Canonical's craft tools ship with it, and a decision model (default typesafe/jev-1.13 via OpenRouter) resolves contradictions between facts. Project guide.
• azir-dispatch - herdr dispatcher where Jev picks agent, model, and effort per task, with a JSONL log and live terminal board. Project guide.
• Backdrop AI Provider TypeSafe AI - Backdrop CMS AI module provider for TypeSafe System One decisions and moderation checks. Project guide.
• bank-stmt-dlh (bank statement PDFs to a reconciled ledger, with Jev only for payments rules can't place) - Turns HDFC Bank current-account statement PDFs into a reconciled ledger with category, recurring-payment, cash and runway reports, on a laptop with DuckDB or as a scheduled Databricks job with a dashboard and weekly email. Parsing, reconciliation, rules and reports are plain code and SQL; the one optional AI step sends leftover payments to Jev behind a 0.9 confidence gate. Project guide.
• banterarena-engine - Open Go scoring engine behind Banter Arena: pure-function banter and country scoring, attribution checks, and two-tier reply sentiment where free rules settle obvious replies and an optional TypeSafe Jev adapter judges whether a reply laughs, concedes or calls the joke weak; labelled yardstick in tests. Project guide.
• barmkin-mod - Claude Code mods security layer: secret redaction on every tool call, untrusted-content taint tracking with an injection screen and a Rule-of-Two egress gate, MCP server allowlists and post-write semgrep — with a Jev System One client that asks two noul questions per screened result (does it try to instruct the agent? does it contain credentials?) against an OpenRouter- or TypeSafe-style endpoint you configure, falling back to heuristics with a circuit breaker. Project guide.
• basal - Independent open typed-decision models for Polish and English (basal-1.5 mini 1.5B, 4.5B, max 11B) with an inference engine serving a /v1/systemone-style API and a public leaderboard that includes hosted Jev 1.13; not affiliated with TypeSafe. Project guide.
• Baton Agent Orchestrator - Idea-to-reviewed-code orchestrator (jev CLI) where TypeSafe Jev dispatches Claude Code, Codex and OpenRouter agents per step. Project guide.
• Bayes of Our Lives - One Embabel agent with three interchangeable decision services: TypeSafe Jev, OpenAI, and local code. Project guide.
• Beacon (shared, self-improving memory for coding agents, with Jev evaluation) - Memory layer that records sessions across Claude Code, Codex, Cursor and OpenCode and turns useful corrections into shared memory; Jev is the default evaluator that decides whether a session holds a lesson worth keeping. Project guide.
• beam-cli - AgentBeam local monitoring/safety CLI for coding agents, with optional TypeSafe Jev action judging (disabled by default; AGPL-3.0). Project guide.
• bekko-system-one - Small System One decision models (17M–400M) for Yes/No, Choice, and Score (independent open weights). Project guide.
• Better Backlog (filter a repo's GitHub issues in your own words, with Jev deciding each match) - Local web page and CLI that loads a repository's GitHub issues through your existing GitHub CLI login and filters them by a plain-language criterion; each issue gets a decision-model answer with a probability, so you can sort and threshold instead of reading the whole backlog. Project guide.
• Better TypeScript - Go-based TypeScript linter whose semantic mode asks TypeSafe Noul questions over bounded source spans for policy checks. Project guide.
• BetterChat (AstrBot plugin that waits for you to finish typing and uses Jev to decide when to speak in group chats) - AstrBot plugin with two features: in private chats it merges messages sent in a burst and cancels stale replies, and in group chats it asks Jev whether the bot should join in, including filtering @-mentions that need no reply. Chinese documentation. Project guide.
• Beyond CLIP: decision models for industrial image QC - Ongoing research study (R code, data and rendered analysis) comparing Jev and the open-weight Clef-Flash with few-shot CLIP for calibrated defect detection on five industrial image datasets, with a five-stage audit, refine, calibrate and freeze framework. Project guide.
• Bin Replenish Lab (pre-registered test of estimating warehouse bin fill levels from photos, with a Jev comparison) - Companion study to Parcel Lens on real warehouse pick-bin photos (Amazon Bin Image Dataset). A fine-tuned model is compared with Jev (jev-1.14.0, experimental image input) picking empty, low, medium or high on 150 test photos: 99% vs 93% on empty-or-not, and 78% vs 58% accuracy on the four levels, with Jev calling 26 of 45 medium bins low. The repository holds no dataset photos. Project guide.
• bitrate-advisor - Live-stream encoder settings from telemetry/history using TypeSafe Jev (OpenRouter) inside a deterministic safety envelope. Project guide.
• Blackrose - TypeSafe System One decision layer that returns allow/review/block before/after LLM generation. Project guide.
• BLADE - Research code that localizes the decisive error step in failed multi-agent runs using TypeSafe Jev multiple-choice judgments. Project guide.
• blink - Bun CLI that searches a codebase with TypeSafe Jev walkers (Choice over directory entries via @typesafe-ai/sdk); upstream licensing unspecified. Project guide.
• Blueprint Support Ticket Triage - Trida AI open blueprint: calibrated typed decisions (TypeSafe Jev / LLM / offline) for support routing and priority, stopping at a human gate. Project guide.
• Bongard - Encoder-decoder System One model returning calibrated probabilities (not text); independent of TypeSafe-hosted Jev. Project guide.
• Bookmark Finder - Local CLI and HTTP API that searches your GitHub stars and X bookmarks: SQLite + BM25 (and optional embeddings) retrieve candidates, then Jev optionally reranks them with a 0-3 relevance score and confidence. Project guide.
• bouncer (Faycall1l) - Programmable live-chat moderation in one ~30ms System One pass using Laya (Jev-compatible) with per-server rules. Project guide.
• BoundedCode - 8 GB GPU coding agent: local generation, required TypeSafe Jev decision plane, and Go-supervisor verification before accept. Project guide.
• BranchOne by Necati Demir (open System One decision model that scores supplied choices from text and an optional image) - Open decision model in the Jev category: send a situation, a question and 2–150 candidate descriptions (optionally one image) and it returns a probability per candidate and the selected ID. Built on Qwen3.5-0.8B-Base with LoRA adapters and a small scoring head shared across candidates; the release includes calibration and inference code, and a separate GitHub repo has data prep and training. Not TypeSafe's Jev. Project guide.
• Brave Jev MCP - Fork of the Brave Search MCP Server whose brave_llm_context results are screened by TypeSafe Jev against the purpose of the search, removing pages and passages that match keywords without answering the question before they reach the AI. Project guide.
• Brier - Local Portuguese-focused typed choice/score/noul decisions in Jev API form (Qwen3-1.7B + LoRA on MLX); independent of hosted TypeSafe Jev. Project guide.
• BrighTO Router - Self-hosted Rust LLM gateway with OpenAI/Anthropic-compatible routes plus System One / Jev / DJEV / Laya decision routing, load balancing, team keys, and usage budgets. Project guide.
• brojev - Go browser agent for Claude Code with its own CDP engine, where Jev (via OpenRouter) picks each action and target during a delegated browser sub-task. Project guide.
• Browser runs for Pi Pocket (goal-driven browser runs decided by Jev) - Pi Pocket extension (0.22.0+) that hands the browser one goal and the values it may type; a run model, Jev through OpenRouter or the session's own model, picks each step's operation, target and value until the goal is visibly done or a clear stop reason applies. Project guide.
• browser-laya - Runs the open Laya System One decision model entirely in a browser tab (ONNX on WebGPU or WASM): an npm library @wexare/laya-web and a hosted playground answer typed choice / score / noul questions in the Jev request shape with no server or API key once the ~900 MB weights are cached. Project guide.
• btx-skill-jev-judge - Claude Code plugin that hands repeated judgments to TypeSafe Jev from a tested script (redaction, rate limits). Project guide.
• Bud Decision Studio - Cross-platform desktop app and localhost server for open Jev-like System One models (Laya, Kev, Lev, Jev-Omni, …) that speaks TypeSafe's /v1/systemone format so Jev SDK code can switch base URL; independent of hosted Jev; no LICENSE file. Project guide.
• Building with Decision Models - Unofficial agent skill that teaches coding agents to design for decision models (TypeSafe Jev, Cloudflare Clef, pplx-decider, OpenAI Decisions, Databricks ai_decide, Ollama models and more): typed questions, calibrated confidence, per-provider reference files and a prior-art library of 450+ community projects sorted by pattern. Project guide.
• Business Insights v2 (bi2) - CLI and MCP server that finds a business's name, logo, merchant category code, head-office address, phone, email and social links from its website, rules first and Jev and Gemini last, with a status, confidence and evidence for every answer. Project guide.
• byo-kit (bring-your-own-data and bring-your-own-key for static demos, with a Jev Decisions helper) - Two dependency-free Web Components for static demos: a drop zone that reads JSON, NDJSON or CSV in the browser, and a masked key field that keeps the visitor's API key in memory or session storage; includes a network-call counter, CSV export that blocks formula injection, an optional CORS relay, and an openrouterDecide helper that calls Jev through OpenRouter's Decisions API. Project guide.
• cai - Rust CLI for AI tasks across many providers that adds TypeSafe Jev subcommands (cai noul, choice, score, jev) for typed questions about text piped on stdin. Project guide.
• Cairn Jev Lab - Experimental memory-admission lab: TypeSafe Jev judges source-backed proposed memories; policy recommends save/skip/defer with recorded demos. Project guide.
• cake-safety-judge (command-safety gate for the Cake agent, with an optional Jev fast-approval cascade) - Standalone Go gate for the Cake coding agent's cake-gate/1 protocol: it judges each proposed shell command with an LLM rubric and strict verdict parsing, and in cascade mode a Jev probability of at least 0.85 approves clearly safe commands quickly while everything else falls back to the judge; offline fake and record providers for tests. Project guide.
• Calendar Curator (Home Assistant) - Home Assistant custom integration that reads events from your calendars, asks Jev whether each one matches a criteria prompt, and mirrors the matches into a target calendar, with caching, reconciliation and a deletion cap. Project guide.
• call-eval - Post-call customer-experience evaluation for contact-centre transcripts: Jev (or an offline rule extractor) pulls typed facts with evidence spans, deterministic code turns them into satisfaction and sentiment scores, and a calibration layer anchors scores to surveys — all on synthetic data with a sealed holdout. Project guide.
• call-pele-answer (agent skill turning answers into interactive pages with Jev routing) - Open-source agent skill for Claude Code, Codex CLI, Cursor and OpenCode that turns a long AI answer into a single offline interactive HTML page, splitting it into blocks and choosing a diagram, table or text form for each; Jev routes block forms in about 200 ms. Project guide.
• Called It! (grade someone's annual predictions at strict, standard and lenient bars, with Jev as the grader) - Claude Code skill that finds a person's or company's annual predictions, turns each into three bars (strict, standard, lenient) before gathering evidence, then grades every claim and bar against the year's public evidence with Jev, committing every receipt so anyone can replay the scorecard; a gallery site shows graded subjects. Project guide.
• Callpoint - Incident decision-support demo where TypeSafe Jev judges live facts and deterministic code makes the call. Project guide.
• Camunda Jev AI Decision Connector - Camunda 8 connector for TypeSafe Jev: typed noul/choice/score decisions in BPMN over lists of any size. Project guide.
• Can you fool Jev? - 1,000 English trick questions sent as identical /v1/systemone requests to Jev, Nimble 9B, tev1 4B and Strands Decider 2B, with the dataset, every request body, per-model answers, a runner for any /v1/systemone endpoint and a video walkthrough. Project guide.
• Can You Trick an AI Shopper? (study with a Jev trick shield and router) - Preregistered study of store tricks on AI shopping agents (47,215 decisions across five models); in phase 2 Jev labels every listing line to strip sales pressure and AI-directed text, and routes easy shelves to a small model and hard ones to a large one. Project guide.
• Candidate Experience Feedback Benchmark - Compare TypeSafe Jev vs LLMs on 60 synthetic candidate-experience reviews (four typed classifications). Project guide.
• Canny - Claude Code/Codex warden: append-only ledger and deterministic done-gate; optional TypeSafe Jev advises but never alone blocks. Project guide.
• Canopy-Jev - Tree-based decision model on Qwen with shared context and isolated branches (independent System One–style research). Project guide.
• CanvasTTY Assistant - CanvasTTY plugin: System One (Jev/Laya/Eikos) command review, launch triage, and orchestrator advice (requires plugin API v2). Project guide.
• cap-jev (SAP CAP plugin) - Plugin for SAP CAP Node.js applications that adds an internal system-one service for choice, score and yes/no questions against one text or JSON state, defaulting to TypeSafe Jev and working with other Jev-compatible System One endpoints. Project guide.
• Captain Code - One terminal that routes tasks across Claude Code, Codex, Cursor, and opencode models, with an optional Jev decision leg for triage, an action gate, and supervision. Project guide.
• catherd - Herds coding agents from your Claude Code session: Claude plans and verifies, Codex/opencode write code, TypeSafe Jev picks the model. Project guide.
• chamoji (changelogs from gitmoji or Conventional Commits, with Jev sorting the commits rules can't place) - CLI and library that reads git history, parses gitmoji, Conventional Commits or plain messages, collapses dependency noise, groups commits into changelog sections, suggests the next semver version and renders Markdown, HTML or JSON. With --ai, commits no rule can place are sorted by Jev. Project guide.
• changelog-bot - AI changelog CLI and GitHub Action with an experimental --why-engine jev: Jev checks each PR-description candidate for an explicit reason that applies to the change, and the strongest accepted text is quoted without paraphrase. Project guide.
• chatwoot-workers - Cloudflare Workers for Chatwoot: Discord relay plus ticket router that triages new tickets with TypeSafe Jev. Project guide.
• Chicken Bench (decision APIs playing Crossy Road) - Playful benchmark of how far, fast and cheaply decision APIs can play Crossy Road, comparing jev-latest and jev-preview with GPT and Clef models on score, latency, tokens and cost per game, with best-run videos. Project guide.
• Chinese-Jev - Chinese-language System One–style typed-decision training and eval pipeline (CJ-Bench/CJ-Dataset); pretrained weights planned, not bundled (independent of hosted TypeSafe Jev). Project guide.
• chinese-workflow-decision-bench - Feishu-style Chinese message classification bench: 64 frozen scenarios, Choice/four-Noul workflows, published TypeSafe Jev vs Laya results. Project guide.
• Chirp (Tern) - Tiny robot voice for the Tern terminal: when a coding agent finishes or waits for you, Jev reads the end of its last message and answers six typed questions (mood, voice, pitch, pace, energy, and whether you're being asked something), and a small Luau synthesizer turns the probabilities into one short beep; without a key it falls back to a local reading. Inspired by Yohei Nakajima's Beep Jev. Project guide.
• Choice Four - Four-in-a-row against Jev: the rules engine finds legal moves and tactical facts, and every Jev turn is a structured Choice over the remaining columns, with server-authoritative matches, Discord sign-in, leaderboards and a hash-linked move journal. Project guide.
• Choice Service (AstrBot plugin for typed choice and yes/no judgements over System One) - AstrBot plugin that gives the bot and other plugins a shared judgement service: it calls Jev or another System One-compatible model with given options or yes/no propositions and returns structured results for proactive replies, intent routing and action choice, with built-in TypeSafe, OpenRouter and Vercel AI Gateway presets; Chinese documentation. Project guide.
• Chon (open Gemma-4-31B LoRA decision model and server that speaks the System One wire format) - Serving code, Docker image and guide for Chon-1.1-31B, an Apache-2.0 LoRA adapter on Gemma-4-31B that answers typed choice, yes/no and score questions with a calibrated probability for every option in one forward pass. The HTTP server speaks TypeSafe's /v1/systemone format so JevBench's typesafe adapter and other System One clients work unchanged. Project guide.
• ci-triage (samiulhuda360) - GitHub Action that classifies failed CI with TypeSafe Jev: re-runs flaky/infra once and comments real failures on the PR (distinct from criguex/jev-ci-triage). Project guide.
• Cite - Elixir library that finds what you describe in a document — failures in a log, hallucinations in a transcript — by gathering candidate passages with fixed rules in code and having a decision model, via the bundled Cite.Provider.TypeSafe (Jev), judge each one; questions are written in a Spark DSL. Project guide.
• ck-browser (plain-language browser checks) - Runs browser checks written as plain-language steps in a real Chrome, with Jev picking each element from a DOM snapshot and deciding whether the page has settled, and writes per-step probabilities, timings, cost and screenshots. Project guide.
• Claim Laundering (paper code: decision models trust an agent's record over reality) - Paper code from the University of Southern California showing that fast decision models, hosted Jev among them, follow an agent's own record of success when an action silently failed, across three domains, and that a linear probe on a middle layer (EviGate) can veto those decisions in open models. Project guide.
• Claims, Choices and Labels (jev-eval) - Audit of TypeSafe Jev and five generative models on Taiwan-status claims, forced choices and labels; draft preprint, item explorer and replays. Project guide.
• ClaimSignal (insurance claim triage with Jev on Cloudflare Workers AI) - Insurance claim triage service for Cloudflare Workers: Jev answers claim type, fraud-indicator score and needs-review in one call, a routing policy auto-resolves, falls back to an LLM or sends the claim to a person, and a benchmark harness compares Jev with an LLM. Project guide.
• Classificador de observações (browser tool that sorts free-text production notes into categories, with Jev as one provider) - Browser prototype that classifies free-text production notes from a spreadsheet into configurable categories, with validated output, parallel batches, fallback when a batch fails and a live cost cap. Groq and OpenRouter LLMs classify in batches; with Jev each row becomes one yes/no question per category and a priority order picks the main category. README in Portuguese. Project guide.
• Classification without fine-tuning (Jev vs. BERT vs. LLMs on 30,000 product names) - Companion repo for a talk comparing six ways to sort ~30,000 Croatian retail product names into 10 categories, from TF-IDF and fine-tuned BERTić to OpenAI Decisions, Claude Haiku, Qwen and Jev, on the same test items for accuracy, latency and cost, including learning curves and unseen-brand results. Project guide.
• Classifier Arena - Three small browser games (a real-time fighter, a spam and smishing lane game, and a reverse-dictionary clue game) where Jev and OpenAI Decisions play head to head on one shared classifier core, with every match logged and replayable. Project guide.
• classifier.dev - Zero-shot text classification over plain HTTP, a CLI, and MCP: a Cloudflare Worker packs inputs into TypeSafe Jev requests and returns a label with calibrated confidence. Project guide.
• Claude Subagent Router - Claude Code plugin with seven subagent workers, a review skill and a Codex CLI bridge; optional model routing sends Agent-call briefs to Jev and switches model when it is confident. Project guide.
• Claude x Jev - Claude Code skill/commands that use TypeSafe Jev on OpenRouter for classify/route/gate so Claude only deep-reads unsure items. Project guide.
• claude-auto-effort - Claude Code mod where TypeSafe Jev rates, in about 200 ms before each turn, how much careful reasoning a prompt needs (0–3) and sets the effort level with smoothing; Claude Code port of pi-auto-effort sharing its questions, policy and settings file. Project guide.
• claude-code-jev - Claude Code PreToolUse permission gate: TypeSafe Jev via OpenRouter Decisions classifies allow/block/ask, with fixture latency/cost benchmarks. Project guide.
• claude-jev (darwintechlab) - Claude Code plugin/MCP for live TypeSafe Jev Choice/Noul/Score tools with confidence auto/escalate labels. Project guide.
• claude-jev-advisor - Unofficial Claude Code command hooks using TypeSafe Jev: a context helper that suggests /compact or /clear at natural stopping points, and (on Windows) an rm helper that asks before real files are deleted but lets Jev clear deletes of session-made or git-ignored files; README lists exactly what is sent to Jev. Project guide.
• claude-jev-funnel - Claude Code plugin + zero-dep CLI: bulk TypeSafe Jev judgments; only the uncertain band goes to Claude/human. Project guide.
• claude-jev-router - Claude Code plugin (Indonesian README) that asks Jev three questions per prompt to pick the cheapest sufficient Claude tier and thinking effort, with a prompt-cache guard. Project guide.
• claude-model-guard - Claude Code hooks plus a CLI picker where TypeSafe Jev chooses the model tier (Haiku, Sonnet, Opus or a long-form tier) for every prompt and can block a prompt sent to the wrong model when it confidently disagrees; tier criteria are editable and there are no dependencies beyond Node's fetch. Project guide.
• claude-model-router by prajeshgpat (local proxy that picks the cheapest sufficient Claude model per turn from Jev's answers) - Runs Claude Code through a loopback proxy: on each new user turn Jev answers typed questions about the request, a truth table in code turns the answers into a model tier, and the request goes on to Anthropic. Your Claude Code login, tools and sessions work as usual; if Jev fails, it keeps the current tier. Project guide.
• claude-referee - Claude Code referee/guard plugin that asks TypeSafe Jev before risky tool actions. Project guide.
• claude-risk-router - Risk-aware Claude Code model routing: TypeSafe Jev answers typed risk questions; rules pick Opus/Sonnet/Haiku. Project guide.
• claude-router (alexei-led) - Claude Code Mod that sets model and effort per turn, advised by TypeSafe Jev (default) or another classifier; no local proxy. Project guide.
• ClauJev (Jev checks for Claude Code) - Claude Code plugin that asks Jev short typed questions on six hooks to triage each request, pick the reasoning effort, guard destructive or credential-touching commands, review edits against plain-language rules and block an unfinished "done". Project guide.
• Clay JEV People Ranker - Clay CLI people search plus TypeSafe Jev Choice/Noul qualification for semantic role fit (bundled operating-founder example). Project guide.
• clear-head - Claude Code Stop hook that judges answer claims against session tool evidence with TypeSafe Jev. Project guide.
• clef-mcp - MCP server (Japanese docs) that lets Claude Code ask Cloudflare Clef — a Jev / System One API-compatible decision model hosted on Ollama — for typed decisions: one decide(state, questions, images?) tool wraps Ollama's POST /v1/systemone with noul / choice / score questions (up to 64) and optional local images, plus a health() check. Project guide.
• clef-webcam - Runs Cloudflare's open-weight clef-flash decision model locally on an Apple Silicon Mac and turns live webcam frames into typed decisions (noul / choice / score) using the same question schema as Jev / System One, with Metal kernels that cut a webcam frame with three questions from ~840 ms to ~250 ms on the author's M5 Max. Project guide.
• cleffa - Native C+Metal Apple Silicon engine for Cloudflare Clef/Clef-Flash with Jev/System One request/response API (MIT; independent of Cloudflare). Project guide.
• clickhouse-jev - Ask ClickHouse rows plain-language questions through executable UDFs backed by TypeSafe Jev (inspired by pg-jev). Project guide.
• CLM (Contrastive Language Models) - Independent open System One model (CLM-8B: Qwen3-8B embeddings plus a small contrastive state/action head) served behind a TypeSafe-compatible POST /v1/systemone, so Jev-style requests replay unchanged; not affiliated with TypeSafe. Project guide.
• CloakBrowser-Agent - Stealth browser agent where TypeSafe Jev decides each step and CloakBrowser executes (MCP/CLI/Python). Project guide.
• clonemap (finds helpers many Python apps rewrite by hand, labelled with Jev) - Research probe that clones a few thousand Django and LLM apps, groups functions that do the same thing with different code, compares them with a control set and lists candidates for a missing library; Jev labels the top clusters and checks proposed libraries. Project guide.
• Clutch (automatic Haiku, Sonnet or Opus choice for each Claude Code message, with Jev deciding) - Claude Code plugin that picks the model and thinking effort for every message you send: a "thanks!" runs on Haiku, a hard bug on Opus with deep thinking, everything else on Sonnet. A short Haiku session brief gives the decider context, and per-project floors and ceilings keep production code above a tier or scratch repos below one. The decider is Jev, or a local llama.cpp /v1/systemone server. Project guide.
• cmd-mod-jev-guard (Jev auto mode for Command Code) - Command Code mod that risk-scores every tool call with Jev and answers deny, ask or allow, denies calls that follow instructions planted in content the agent read, scans tool results for prompt injection and sweeps instruction files. Project guide.
• cmd-mod-jev-nudge - Command Code mod: TypeSafe Jev nudges the agent to continue when stop leaves unfinished work. Project guide.
• Coach (desktop app for defining credentials and testing which of your agents can earn them, with an optional Jev Mode) - Local desktop app: define a credential as learning targets and checkable assignments, register the agent setups you run (harness, model, effort), and the app writes test prompts and handoff notes, records results as plain Markdown and JSON, and shows which agent to reach for. Optional Jev Mode, off by default, sends text to TypeSafe only when you ask. Early prototype (Tauri, Mac and Linux). Project guide.
• Cobalt Cockpit - Mission control for Claude Code (HUD, verification gates, Run Ledger) with an optional per-session TypeSafe Jev advisory router. Project guide.
• Code Quality (smixs) - Advisory code-quality report and red-flag hooks for Claude Code, Codex, pi, omp, OpenCode and Grok, with an optional TypeSafe Jev classifier that asks calibrated yes/no questions about added test and source hunks (never blocks). Project guide.
• codec-swarm - Local single-user LangGraph orchestrator that takes a Jira ticket to a reviewed local PR with a swarm of Claude Code agents (specifier, architect, coder, reviewer, hardener, QA, judge) in per-repo worktrees; with Jev enabled, every approved Gherkin scenario and every uncovered ticket criterion is scored in the judge step, and risky actions (push, migrations, terraform apply) always ask you. Project guide.
• codeforces-tag-prediction - Leak-free multi-label Codeforces tag benchmark comparing TypeSafe Jev to zero-shot LLMs and encoders. Project guide.
• codegraph-jev - Research log: can a code graph plus TypeSafe Jev match a coding agent at finding and ordering functions? Project guide.
• Codex Decision (trim repetitive Codex tool output with OpenAI Decisions or Jev) - VS Code extension and native hooks for Codex that select repetitive lines for removal from local tool output while keeping errors and useful evidence visible, with a link to the full original; OpenAI Decisions is the default and TypeSafe Jev or a custom endpoint can be chosen instead. Project guide.
• Codex Jev integration - Standard-library Python MCP server for Codex CLI and App providing advisory role and resource selection with Jev: six tools (jev_status, jev_route, jev_decide, jev_select_resources, and two outcome reporters) send task goal, constraints and eligible candidates as choice questions to jev-latest, with deterministic review gates, local validation and confidence fallback; it never executes, launches, switches models or approves. Project guide.
• Codex Jev Preflight - Fail-open Codex UserPromptSubmit hook: TypeSafe Jev injects advisory task_type/complexity/risk/execution_mode. Project guide.
• Codex Jev Router (suenot) - Uses Jev Choice and Noul to choose Codex subagent model and reasoning effort, with local confidence gates and a Sol fallback. Project guide.
• Codex-Jev - VS Code Codex plugin that shortens noisy tool results with Jev-gated evidence selection before Codex reads them. Project guide.
• codex-jev-router - Local Codex CLI bridge: TypeSafe Jev selects model and reasoning effort per turn via a loopback Responses proxy (fail-open without a key). Project guide.
• codex-route-observer - Local proxy that observes Codex/OpenAI-style coding-agent traffic and uses TypeSafe Jev to categorize current request usage for a dashboard of routes, models, and spend-ish signals. Project guide.
• codex-router - Codex CLI model router that picks the model and reasoning effort for each turn (optionally via a System One classifier such as Jev) and shows the choice in Codex's footer. Project guide.
• codex-triage - Local Codex task triage with human-reviewed archiving and optional TypeSafe Jev analysis. Project guide.
• Codify (cg) - Agent workflow CLI (cg) with a spec engine, code graph, and memory; cg jev asks TypeSafe Jev for typed judgments such as memory classification, flaky-failure triage, and finding ranking. Project guide.
• Coding Router Jev - TypeScript/Bun launchers codex-jev and claude-jev plus a Pi jev/auto virtual model that route each coding-agent user turn through TypeSafe Jev to pick a model tier (Fast, Balanced, Strong, Long) and reasoning effort, with confidence-capped local policy. Project guide.
• cog-typesafe - CogCloud provider Cog: TypeSafe Jev as an explicit versioned system-one/decisions binding. Project guide.
• College Event Moderator - Benchmark and local playground that screens college chapter event proposals with Jev or a local s1ced or Vega model: five shared questions per event, expected answers, and a report of accuracy, confidence, latency and cost. Project guide.
• ComfyUI-ScriptFlow - ComfyUI safe script node with TypeSafe Jev yes/no/choice/score decisions (local GGUF LLM fallback) for workflow branching. Project guide.
• compact-adviser - Agent plugin that asks TypeSafe Jev whether the session is at a safe /compact boundary (hint or optional auto-compact). Project guide.
• Compass (ring29 labs) - zsh helper that keeps your native completions and adds a plain-English request: type a CLI prefix plus what you want (e.g. aws ec2 and 'list all instances in table format') and Compass gathers candidates from completions, shell history and local help, then asks TypeSafe Jev to rank them; the chosen command is inserted for review, never executed. Project guide.
• Competitor Monitor with Jev (Mastra template) - Official Mastra template that watches competitor pricing, changelog, and docs pages, diffs them against saved history, has TypeSafe Jev classify each change, and applies a policy to ignore, record, request review, or alert, with an LLM explaining results. Project guide.
• Condere (open-source spin-off analyst for a code repository, with Jev judging every run) - Developer preview that inspects a software repository, gathers outside evidence, compares commercialization routes, and returns a pursue-or-stop decision with linked reasoning and a build plan a coding agent can read. Jev and a second model judge published results on every run, with an optional paid Opus check. Project guide.
• Conjevture - Exact logic for uncertain answers: TypeScript library combining Boolean rules with Jev Noul/Choice probabilities (zero runtime deps; MIT). Project guide.
• Conjure (decision model art-directs Three.js scenes) - Type an idea or mood and a decision model picks worlds, things, looks and directions from a catalogue of procedural Three.js components, then builds and prints the scene in seconds; works with Jev directly or other System One hosts. Project guide.
• Connector for TypeSafe Jev (WordPress) - WordPress plugin that gives themes and plugins a typed way to ask Jev about content and branch on a probability, a named choice or a score with confidence, with the key managed by WordPress core's Connectors API; a companion plugin backs comment moderation with it. Project guide.
• content-evals (agent skill) - Agent skill that breaks a content requirement into atomic yes/no questions, builds a YAML scoring profile, scores a document with Jev and turns low scores into concrete edits. Project guide.
• context-budget - Context-compaction toolkit that can rank relevance with TypeSafe Jev (/v1/systemone on api.typesafe.ai) to cut most of what agents send upstream while measuring answer quality on a benchmark. Project guide.
• ContextGem (document extraction with LLMs and decision models) - Open-source Python framework for extracting structured data from documents; version 1.0 adds decision models such as Jev, which answer closed questions (yes/no, labels, ratings) with probabilities for each option, alone or in a model group next to LLMs. Project guide.
• contextgem-enconvo (Enconvo plugin for ContextGem extraction that routes closed questions to Jev) - Enconvo plugin that wraps ContextGem for aspect-based retrieval and typed concept extraction over text and documents, using the LLM providers you already have in Enconvo. A decision bridge, on by default, sends booleans, labels and ratings to Jev through Enconvo Cloud for calibrated probabilities, falling back to the LLM when Jev is unavailable. Project guide.
• contextleleo - Move coding-agent sessions between tools with TypeSafe Jev scoring which context the next task needs (built on txcript). Project guide.
• ContextOS (desktop workbench for coding-agent sessions, with Jev-assisted compression of long chats) - Desktop workbench built on Wake that gathers sessions from the coding agents on your machine: browse and full-text search conversations and tool calls, resume or export sessions, track project file changes, manage memories and skills, see token stats, and compress long sessions with rules, Jev judgments and a configurable LLM. Exposes MCP and a CLI. Chinese docs. Project guide.
• Copilot Studio × Jev - Copilot Studio/Power Platform: MCP gates Azure AI Search passages with TypeSafe Jev yes/no questions; plus a TypeSafe custom connector (MIT). Project guide.
• corp-jev-skill - Claude Code /corp-jev skill that runs a corporation of review departments with TypeSafe Jev second opinions via OpenRouter. Project guide.
• could-jev - Claude skill that gives a ten-line gut-check on whether TypeSafe Jev could help with what you are building: verdict, matching pattern from 31 tagged Jev patterns, a state/questions sketch, closest prior, pitfalls and the cheapest test; it does not call Jev itself. Project guide.
• CourtVision (basketball computer vision with a Jev decision layer) - Real-time basketball pipeline that turns single-camera video into verified stats, then uses Jev to classify play type and moment stakes, rank coach cards and guard Claude's narration lines, with an operator dashboard over WebSockets. Project guide.
• cpa-plugin-typesafe - CLIProxyAPI plugin exposing TypeSafe Jev typed API with OpenAI compatibility and private settings. Project guide.
• craft-cli - Unofficial local-first CLI and TypeScript library for Craft Docs, with an agent skills library stored in Craft; craft lib pick --jev can use TypeSafe Jev to choose which skills fit a request. Project guide.
• CRAG Gate Benchmark (Jev vs. GPT-4.1 Mini as the filtering gate in corrective RAG) - Audited benchmark asking whether Jev can replace GPT-4.1 Mini as the product-filtering and routing gate in a corrective RAG workflow on Amazon ESCI: 800 calibration and 2,000 held-out queries with a shared BGE reranker and real OpenRouter calls; it publishes the failed acceptance condition as well as the precision, latency and cost gains. Project guide.
• CrayonCli (.NET CLI that asks Jev which of 8 crayon colors an image is closest to) - Small .NET 10 prototype that sends an image to Jev and prints which of eight classic crayon colors it is closest to, with a probability for each, plus a script that checks Jev against labeled test images. Project guide.
• crea-branding-score (0-100 branding score for video ads, with Jev checking each frame for the logo) - Scores how much a set of video ads shows a brand: speech-to-text finds brand mentions in the voice-over, and Jev answers typed questions on each frame (is the logo visible, compared with your reference logos) through OpenRouter's Decisions API; outputs a per-ad score with a breakdown, group averages and an HTML report. Project guide.
• Creative Decision Making (cdm) - Paper and Python library that combine calibrated decision-model judgments (Jev and peers) with decision theory: calibrate, pool, gate unknowns to verify, and pick options plus bounded bets on new ideas, with a benchmark on 237 real Jev judgments. Project guide.
• creator-teardown - Claude Code/Cowork skill that tears down a creator account, labels posts/sentences with TypeSafe Jev, and drafts rebuilt variants for your brand. Project guide.
• CronSight (scheduled inspection agent platform with optional Jev anomaly triage and human review) - MiniClaw-based platform for scheduled health checks: a read-only runner checks error rate, latency, CPU, backlog, certificates and data freshness from a workspace JSON snapshot. An optional path sends anonymised metrics to Jev as a Choice of anomaly, normal or uncertain; low-confidence results go to a human review screen, and alerts leave through a separate outbox. README and docs in Chinese. Project guide.
• Crosstalk - Lab for finding and fixing look-alike option confusion in typed-decision routers; live mode runs against TypeSafe Jev. Project guide.
• Cruise (Claude Code context, progress and effort plugin) - Claude Code plugin from NTConsult with context warnings, a progress band above the prompt and an optional effort judge that picks a lower reasoning effort for small requests, where the judge can be Haiku, Jev on TypeSafe or a local Laya model. Project guide.
• Cua jev-use - Composes a bounded Jev chooser with Cua Driver and an independently verified browser fixture. Project guide.
• CUA-JEV (ZJU-REAL) - Jev-powered computer-use framework: adapters propose legal actions; Jev chooses action and channel; code guards and verifies. Project guide.
• Customs Authority (Claude Code toolkit that finds where Jev fits in your app and ships a confidence-gated hybrid) - Claude Code plugin that walks a project's decision points, triages each one as code, Jev or LLM, helps write typed questions, proves them against an answer key, and drops in a gated hybrid router that sends low-confidence items to a person or a larger model. Includes a dependency-light Python Jev client and TypeScript templates. Project guide.
• Custos (Rust behavioral security engine and firewall bouncer that asks Jev to judge suspicious IPs) - Single-binary Rust engine inspired by CrowdSec: it tails web access logs or runs as Axum middleware, drops obvious noise and floods for free, strips prompt-injection text from user agents and URLs, and sends batches of up to 16 suspicious IP windows to Jev in one call for a ban probability and a threat type. Bans go to an HTTP bouncer for Nginx, Caddy or Traefik and to OS firewalls (pf, nftables, ipset, netsh), with an optional gRPC hub for sharing signals between nodes. Project guide.
• CV Assistant + Jobpilot (job-search skills and a local dashboard that scores jobs against your evidence with Jev) - Job-search kit: seven skills for Claude Code, Claude.ai, ChatGPT and Codex that turn your real experience into an evidence-graded skills pool and a tailored ATS-friendly CV per ad, plus Jobpilot, a local Windows dashboard that collects jobs from public boards (Greenhouse, Lever, Ashby, peviitor.ro), scores them with Jev and can fill and apply to the best ones. Auto-submit is off by default. Romanian docs with an English summary. Project guide.
• Cygnet recipe - Independent recipe that answers Jev-format typed decisions (choice, noul, score) with frozen Gemma-4-12B-it on unmodified vLLM via a one-token option-letter readout and a /v1/systemone shim; not affiliated with TypeSafe. Project guide.
• Dadvar (open training data, training and evaluation for System One decision models) - 36 public Hugging Face datasets converted into TypeSafe System One rows (state, questions, gold answers), code to fine-tune a calibrated decision model from a pretrained encoder, and a benchmark runner that scores any model, including TypeSafe's hosted Jev, on accuracy, calibration and latency over the same rows, with diagnostics for question reading, references into the state and option order. Project guide.
• daf-jev - Composable Python toolkit for TypeSafe Jev: question builders, confidence gates, evaluator, calibration, CLI, and optional MCP server. Project guide.
• dailypaper-skills - Agent-skills pipeline (Chinese-language docs) that finds new papers from HuggingFace Daily/Trending and arXiv, uses TypeSafe Jev to score topic relevance against your research interests, then has your agent write reviews and Obsidian notes. Project guide.
• Damocles - VS Code extension and desktop app embedding a multi-provider AI coding agent; with a TypeSafe (or OpenRouter) key, Jev runs its persistent-memory contradiction judge and memory reranking instead of a sub-call LLM. Project guide.
• Data Agent MNIST (ClickHouse) - ClickHouse's harness for building an agentic-analytics benchmark from your own data warehouse, with two optional Jev tools: schema-retrieval ranking (one yes/no per table) and failure-mode labeling of graded runs. Project guide.
• datafusion-jev - DataFusion SQL prompt_jev UDF: typed TypeSafe Jev Noul/Choice/Score over row text (bring your HTTP client). Project guide.
• Dataiku TypeSafe AI plugin - Dataiku's DSS plugin that brings TypeSafe Jev into the LLM Mesh: a Jev Question Set block for Structured Visual Agents, a yes/no Jev guardrail, a Jev reranker and a Jev chat-completion model usable in prompt and Classify text recipes, with usage, cost and traces recorded by the Mesh. Project guide.
• dbt_jev - Classify warehouse SQL values with TypeSafe Jev (or OpenRouter→Jev) from dbt macros on DuckDB and ClickHouse. Project guide.
• Deadline Tetris - Live benchmark where Jev, Clef, Laya, Strands Decider and LLMs play the same Tetris game while pieces keep falling: over four games Jev landed about 119 pieces per game with a 289 ms median answer and won all four, while slower models placed better moves but too late. Project guide.
• decide (CLI for typed System One decisions with named profiles) - TypeScript CLI for yes/no, choice and score decisions where you call a named profile instead of an engine; each profile maps to an engine and pinned model (TypeSafe Jev or OpenAI Decisions), and every result records which ran. Project guide.
• decide (jeffcampbell) - Dependency-free Python 3.11+ CLI and library that asks System One models — hosted TypeSafe Jev by default, or clef-flash locally in Ollama — typed yes/no, choice and score questions about files and text; decide triage tells a coding agent which files are worth reading before they enter its context. Distinct from the listed decide (vsekhar). Project guide.
• Decide (Jev's native decision API as a Go CLI and MCP server) - Go CLI and MCP server that exposes Jev's native question types (noul, choice, score, experimental box, and multi-question systemone) to shell scripts and coding agents, with offline request previews, caller-defined acceptance gates, ordered JSONL batches and one-command MCP and skill install for Claude Code, Codex, Cursor, OpenCode and others. Project guide.
• decide (vsekhar) - CLI to ask TypeSafe Jev typed decisions from the shell, scripts, and agent skills — no application code required (Show HN). Project guide.
• DecideBench - Benchmark of decision models (Jev and open alternatives) and LLMs measuring accuracy, cost per task and latency together on 400 contrastive decisions, with Pareto frontiers instead of a single ranking. Project guide.
• DecideKit - Typed decision policies for TypeScript/Python with Jev via OpenRouter or TypeSafe, offline fixtures, and explicit confidence fallbacks. Project guide.
• Decider.jl - Julia front end for decision models: ask Choice, YesNo and Score questions about some state and get calibrated probabilities keyed by your own @enum or Symbols, with backends for TypeSafe Jev, OpenRouter decision models and the OpenAI Decisions API. Project guide.
• DecideX - Industrial System One decision foundation with TypeSafe Jev provider, confidence gating, and deterministic fallbacks. Project guide.
• DeciForm (structured objects from System One decision models) - Python library that fills Pydantic schemas with System One decision models instead of generated text: enum, bool, bounded-int and optional fields become typed questions with calibrated per-field confidence, with escalation to an LLM only for free text or low confidence. Project guide.
• DeciLabel - GitHub Action that classifies Issues and PRs with TypeSafe Jev (or OpenRouter / OpenAI Decisions) and adds only evidence-backed labels. Project guide.
• Decis - Self-hosted System One–compatible /v1/systemone server with open Laya/kev Docker images (independent of hosted TypeSafe Jev). Project guide.
• Decision Engine for PHP (swisnl) - PHP library for typed, probabilistic Choice / Score / Noul decisions in application code, with TypeSafe Jev as the default engine and OpenAI/Anthropic as alternatives. Project guide.
• Decision Index (apolinario) - Reproduction kit for the Decision Index benchmark suite for typed decision engines (choice/noul): local or one HF Job; not affiliated with TypeSafe. Project guide.
• Decision Layer Benchmark - Rerunnable cost comparison of agent decision layers on OpenAI, Anthropic, Google and TypeSafe Jev. Project guide.
• Decision Model Arena - Arena comparing TypeSafe Jev, OpenAI Decisions, Decision-1 and Laya on the same game states and questions. Project guide.
• Decision Model Rust SDK - Async Rust SDK for decision models served through the System One API (TypeSafe Jev and compatible vendors): typed question builders or #[derive(QuestionSet)], one HTTP/2 connection per client, and an adapter crate for OpenAI/Anthropic/Gemini. Project guide.
• Decision Model Workbench - Local-first workbench and guides for bounded decisions: call hosted decision models (TypeSafe Jev, OpenRouter, Perplexity, Cloudflare) with typed validation and fallbacks, or review data and train a CPU baseline domain model. Project guide.
• Decision models demo (runnable talk demo comparing Jev, Microsoft-Decision-1 and OpenAI Decisions with a generative LLM) - Small zero-dependency Node.js demo for a lightning talk on decision models: send state and typed questions, get a probability per option, and compare a decision model with a generative LLM on the same tasks. One adapter layer supports TypeSafe Jev, Microsoft-Decision-1 on Foundry and the OpenAI Decisions API, and a mock mode runs offline with no keys. Project guide.
• Decision Models Lab - Hands-on lab for decision models: send the same support ticket and typed questions to TypeSafe Jev and Microsoft Decision-1 through OpenRouter, compare answers and confidences, and learn when to hand a case to Claude for careful reasoning. Project guide.
• Decision models on Polish historical gazetteer entries (IHPAN) - Evaluation by the Institute of History, Polish Academy of Sciences: 250 human-labelled fragments of a 19th-century Polish geographical dictionary scored by TypeSafe Jev (via OpenRouter), TEV1 and Basal against human relevance judgments, with data, scripts and raw outputs (Polish docs). Project guide.
• Decision models vs small LLMs (Jev, Kev, Laya and Mammoth-1 against cheap LLMs on agent decisions) - Rerunnable comparison of decision models (Jev 1.14, Kev-4B, Laya, Mammoth-1) with small LLMs on six decisions an agent makes between generative steps: routing, classification, a guardrail, an ordered score and a mixed workflow, scored on accuracy, calibration, format failures, latency and cost. Project guide.
• Decision Tagger (Obsidian) - Obsidian plugin for System-1 note tagging via TypeSafe Jev with multi-key parallel batch classification. Project guide.
• Decision Wall (type a text and watch 200 Jev yes/no judgments light up as a heatmap from one parallel call) - Zero-dependency Node demo: type anything and 200 yes/no questions from one of seven question sets are answered by Jev in a single parallel call, shown as a live heatmap with measured latency. A small server proxies the call so the key never reaches the browser. Project guide.
• decision-agent-lab (measured tool routing and command gating with Jev and other decision models) - Korean-language lab that measures the "LLM generates, decision model judges, code executes" pattern: Jev, OpenAI Decisions, Upstage Solar Decide and local Ollama models are called through one interface on a 40-request tool-routing task and a 1,007-command allow/ask/deny gate, with every call's model version, latency, tokens and cost logged. Project guide.
• decision-cli (Jev-like decision API from local models on Apple Silicon via MLX) - Go CLI (Chinese docs) that deploys small local LLMs on Apple Silicon with MLX and reads candidate-token logits from a single prefill step to return Jev-style class probabilities for batches, text and multimodal input. It can serve an HTTP endpoint compatible with TypeSafe's /v1/systemone, and ships a JevBench public-231 comparison against TypeSafe Jev 1.13.0. Project guide.
• decision-drift (open daily tracker of decision models like Jev: drift, noise, robustness and calibration) - Reproducible tracker that sends a fixed, append-only question suite to every hosted decision model on OpenRouter (TypeSafe Jev, Cloudflare Clef, Perplexity Decider, OpenAI Luna Decisions and others) and reports drift from day 1 beyond each model's noise floor, silent snapshot swaps, sensitivity to option order and paraphrase, accuracy and calibration, latency, errors and cost. Project guide.
• decision-eval - Benchmark harness for zero-shot decision models (choice / noul / score): accuracy, calibration, latency, and cost across suites, with a hosted Jev backend on OpenRouter (typesafe/jev-1.14) alongside local Laya/GLiNER and OpenAI Decisions. Project guide.
• decision-first - Agent skill that tries TypeSafe Jev on bounded-judgment steps first and documents every attempt in a reusable decision lab. Project guide.
• Decision-model referral on medical exams - Code, recorded runs and manuscript for a preprint (Sangzin Ahn, 2026) comparing TypeSafe Jev, OpenAI Decisions and Cloudflare Clef on 435 KorMedMCQA and 1,273 MedQA questions, using each model's selected-option probability to refer its least confident 20% of answers to a stronger model versus random and oracle referral. Project guide.
• decision-model-audit (reproducible field test of Jev and other decision models against small LLMs) - Reproducible comparison of decision models (Jev, GLiNER2.5-Decide, Laya, Clef and thirteen more in part 2) against small LLMs with structured output on fresh September 2026 data: accuracy and calibration, label-wording sensitivity, long inputs, no-fitting-label behaviour and cost per 1,000 decisions, with all raw outputs committed and blog write-ups in English and German. Project guide.
• decision-model-bench (Jev vs. OpenAI Decisions, ai_decide and decider-2b on support chats) - Benchmark comparing typed decision models on the same customer-support chats (ABCD dataset) with the same questions and paired significance tests: hosted Jev, OpenAI's Decisions API and Databricks' ai_decide, plus open decider-2b out of the box, with retrieved examples and fine-tuned with QLoRA, across intent triage and next-action jobs. Project guide.
• decision-models-eval - Independent, pre-registered evaluation of SubQ Snap against TypeSafe Jev 1.13 and 1.14 on Decision Index benchmarks, including a long-context fairness test (whole documents vs chunks), with figures and confidence intervals. Project guide.
• decision-rs (typed Rust decisions across TypeSafe System One, OpenRouter Jev, OpenAI Decisions and Clef) - Rust crate where you declare decision questions once as derived types (Predicate, Choice, Outcome) and choose the backend when constructing the client: TypeSafe System One, OpenRouter's Jev route, OpenAI Decisions or Cloudflare Workers AI Clef, with the same acceptance methods and error categories on each; self-hosted System One endpoints are supported. Project guide.
• decision-server - Cloudflare Worker that answers Jev-format decision calls on free Workers AI Clef models first and falls back to Jev, with REST, a remote MCP server, budgets and a stats dashboard. Project guide.
• decision-shadow - Zero-dependency Python wrapper to shadow-test Jev against an existing LLM classifier, report agreement, tune thresholds, then switch over with fallback. Project guide.
• decision-toolkit - Local-first toolkit from GIA-UH and Syalia for open-weights System One decision models: typed choice, yes/no, and score questions with calibrated probabilities on CPU, as a library, CLI, and REST server that speaks TypeSafe's System One contract. Project guide.
• DecisionBench - Evaluation harness and dashboard comparing TypeSafe Jev (via OpenRouter) with Claude Sonnet 5 on the same structured decision tasks — 100 real invoice_processing cases with five sub-questions each — reporting accuracy vs gold, cost per 1k decisions, p50/p95 latency and a disagreement view; Jev led overall but Claude won duplicate detection (author-reported). Project guide.
• DecisionKit - Provider-independent .NET decision model: typed Choice/Score/Noul questions in the domain, TypeSafe Jev protocol in a separate package, DI and offline test doubles (pre-release). Project guide.
• decisionkit (adityajha2005) - Provider-neutral decision CLI, MCP server and Claude Code plugin with a Jev backend for agents, scripts and CI. Project guide.
• DecisionLint (lint Jev and Luna decision questions in VS Code, CI and MCP) - Linter for the questions you send to decision models: flags mistakes in Jev and OpenAI Luna requests in JSON, JavaScript, TypeScript, Python, Rust and Go as you type, with fixes, a CLI for CI (SARIF and JUnit), an MCP server, and optional Ask commands that have a model review a file. Project guide.
• DecisionOps (calibration and confidence-threshold certification for decision models, with a Jev provider) - Open-source reliability toolkit for decision models: from a decision contract (closed choice set, which choices may run automatically, path costs and miss severities) it measures calibration, certifies a confidence threshold before go-live and monitors it after; providers include hosted Jev, Laya and your own heads. Project guide.
• decisions-lint (lint PHP code against plain-language rules with Jev) - PHP linter for rules static analysers cannot check, such as whether a method belongs to its class or whether a name says what the code does: each rule is a typed question a decision model answers, and each violation comes with the change that resolves it. Project guide.
• decisions-mcp - MCP server that lets coding agents scan large codebases with OpenRouter decision models (Jev by default) to locate bugs from symptoms and classify or filter files cheaply. Project guide.
• decisionsmith - Use and fine-tune System One models (Jev, Laya) on your own data with an LLM as teacher: declare decisions as Pydantic fields and get a harness that calls Jev or a trained local student. Project guide.
• DecisionTools (DeepSeek Harness tool recommendations and search prefetch from a Jev-style Decision API) - Optional DeepSeek Harness (DSH) plugin: a bounded-answer Decision API such as Jev recommends tools before the main model's first request (per-session DecisionBoost) and can preselect and prefetch search sources, while every tool stays available and the main model keeps control. Project guide.
• Decisium (plain-English Gherkin UI tests run in Playwright, with Jev picking the page actions) - Semantic UI testing: plain-English Gherkin steps run in a real Playwright browser, Jev picks among the actions the page actually offers and answers whether each step is done, and code decides pass or fail; record and replay cassettes make CI runs deterministic with no model calls. Project guide.
• Deeds - CLI and Claude Code plugin that reads every commit diff and counts what actually got built (new, deepened, regressed, removed capabilities); TypeSafe Jev answers typed questions per commit and a fixed policy turns them into counts. Project guide.
• Deep Recall - Answer-aware recall over a folder of Markdown notes (Claude Code plugin + Python CLI/MCP) whose pluggable reranker can be TypeSafe Jev, scoring each passage window by the probability that it states the answer. Project guide.
• DeepSeek Harness for VS Code - VS Code extension for DeepSeek Harness with a built-in, optional Laya/Jev decision layer: loop guard, large-output shaping, done-gate, tool pruning, and skill routing. Project guide.
• deepseek-harness-jev (luobosibing2) - DeepSeek Harness (DSH) plugin: TypeSafe Jev for skill/file ranking, supervision, corrections, and workspace approval (features off by default) — distinct from wjw66/deepseek-harness-jev-pre-compaction. Project guide.
• defrag - Research tool for telling a coding agent's user when compacting is safe — at the quiet point between tasks, not when the context is nearly full: it extracts checkpoints from real sessions, lets you label them, and replays them against judges with TypeSafe Jev as the first production judge (plus OpenAI Decisions, rules, a threshold proxy or a local model) to measure which one to ship. Project guide.
• demo-expanso-jev - Expanso Edge × TypeSafe Jev demos: fingerprint routine logs locally, ask Jev only on the remainder, route with hold-on-failure and live boards. Project guide.
• DepthJev - Embodied navigation: RGB→depth/text facts; TypeSafe Jev chooses EB-Navigation actions. Project guide.
• Deqio - Local typed decision API (noul/choice/shared) with swappable engines/backends (Kev, Laya, Open-Jev, and others) and a lightweight UI. Project guide.
• Dernier Abri (Last Shelter) - French student 3D browser defence game where Jev estimates from a text description whether a night silhouette is a zombie or a survivor, and deterministic, tested code decides whether to fire, using Bayesian evidence fusion and a sequential test. Project guide.
• DERVA (AI gateway for regulated teams with a Jev decision layer) - Gateway built on the LiteLLM proxy that sits between AI agents and models for regulated teams: before a request reaches a model, a fast decision layer answers typed questions (prompt injection, client data leaving the firm, regulated advice, which model tier it needs), with Presidio masking, versioned policy, human approval for actions and a hash-chained audit log; the decision layer is vendor-neutral, and the author adopted Jev after it passed all 15 pre-set criteria on held-out data. Project guide.
• DESIGN:OS Generative UI - Streaming generative-UI runtime for agent webviews whose engine includes a TypeSafe Jev adapter (with a Laya-MLX path) for its decision cascade. Project guide.
• deslop - Agent skill/CLI that scores page bodies with TypeSafe Jev probabilities for ads, slop, SEO shape, and derivative content. Project guide.
• Deterministic Text-to-SQL (Jev picks schema slots, sqlglot builds the SQL AST) - Japanese-documented text-to-SQL engine that never lets a model write SQL: Jev classifies the question into dataset, field and metric slots generated from a semantic schema, foreign-key paths come from a NetworkX graph search, and the query is assembled as a sqlglot AST; a mock mode runs offline. Project guide.
• Dev Studio (one-click team of Claude Code sessions, with Jev routing every message between them) - Dock app that opens a team of Claude Code sessions (an Opus architect, a Sonnet main session and Haiku workers, each worker in its own copy of the project) plus a board showing Jev routing every message between them: who sent what, who got it and how sure the router was. Built on Herdr. Project guide.
• dev-decisions - Verification and calibration layer for AI-aided development (v0.3.0): git hooks block secrets and classify diffs, an optional global ZCode hook blocks agent-run commits/pushes that fail the checks, and a CLI runs plan-contract gates and outcome grading — routing classification through a roster of decision models where TypeSafe Jev (jev-1.13.0) handles contract gates and fan-out next to GLiDE, local GLiNER2 and others. Project guide.
• dgem (DiffusionGemma as a zero-shot decision model with a /v1/systemone-compatible API) - Turns DiffusionGemma 26B-A4B into a decision engine: write a decision as a template of typed questions (boolean, choice, score) and get per-option probabilities and a hesitation score from one pass, with hesitant cases handed to Gemini. Comes as a CLI, an HTTP gateway including /v1/systemone, an MCP server and a Decision Studio web app, plus benchmark harnesses and deployment guides for Apple Silicon, NVIDIA, Cloud Run and Vertex AI. Not an officially supported Google product, and not TypeSafe's Jev. Project guide.
• Did the News Move the Price? (Dhaka Stock Exchange event study, with Jev scoring the news) - Code, data and manuscript for an event study of 70 company-specific news events on the Dhaka Stock Exchange with a same-firm placebo: Jev scored each headline and article (is it about the company, is it good or bad news), and the finding is no detectable average price reaction, with only a weak negative asymmetry for bad news. The public package reproduces the numbers offline. Project guide.
• Discern - Effect Decision/DecisionModel patterns, policies, and procedures with an optional TypeSafe Jev provider (@doeixd/discern). Project guide.
• discoprint - Classify an artist discography for theme, mood, and lyrical complexity with TypeSafe Jev and an Ink terminal dashboard. Project guide.
• Discord Policy Checker - Experimental Chrome extension that checks outgoing Discord messages against your policies with Jev, with optional Gemini rewriting. Project guide.
• discourse-chatbot - The original Discourse AI Chatbot plugin; optional System One settings (default model jev-latest) let Jev decide whether a question belongs to the community and which tools the bot should use. Project guide.
• Distill - Lightweight coding-agent TUI with TypeSafe Jev (or OpenRouter decisions) for model/effort routing, utility tasks, and retention. Project guide.
• DMLab TypeSafe Core (Jev for Magento 2) - Magento 2 module that connects a store to TypeSafe's Jev with one encrypted key, a pinned model version, a monthly token budget, a response cache, retries and a circuit breaker, shared by DMLab AI modules such as Catalog Enrich, which fills empty product filter attributes from descriptions. Project guide.
• doc-router - Routes PDF pages between local text extraction and OCR using optional Jev judgments, with a Rust CLI and Python bindings. Project guide.
• DocJev - Document classification and splitting with TypeSafe Jev over LiteParse (optional LlamaParse) page text. Project guide.
• Docket (pg-jev Phase 0 benchmark) - Design and reproducible Phase 0 benchmark for durable, versioned, calibrated AI decisions in Postgres: measures jev-1.13 through pg-jev on 2,000 real complaint narratives for calibration, determinism and context effects. Project guide.
• docseek - Goal-driven document crawler where a relevance judge (Jev is one built-in option) scores links, ranks pages, flags pages that hide documents, and picks filter values. Project guide.
• Document classifier (financial document types) - Node app and benchmark that names a financial document's type (10-K, proxy statement, fund fact sheet, trade confirmation and 17 more) with one Jev Choice call, compared on 43 synthetic documents against frontier and small LLMs, a bi-encoder and keyword rules. Project guide.
• Doom in Claude Code (Jev pilots) - Claude Code plugin that runs ViZDoom with Freedoom in a terminal pane: Jev looks at the game screen and picks the strategy about 2.5 times a second, Claude Haiku narrates in Spanish or English, and Python reflexes and an A* navigator handle aiming and the map. Project guide.
• dopp - Point Jev-shaped decision calls at one URL, keep every request, train a small model you own, run hosted/offline/browser—independent path. Project guide.
• doublestrain - Grades dependency advisories into SSVC Act, Attend or Track using whether the vulnerable code is reachable in your app, with a fixed decision tree and model judges (including an optional Jev classifier) held to measured acceptance floors. Project guide.
• Dowse (semantic find for Safari) - macOS app and Safari extension that opens from the MacBook notch: describe a passage you remember, and Jev finds and highlights the matching text on the page so you can step through matches. Project guide.
• driftnet (drop links in chat and get them sorted, explained and searchable, with Jev as the judge) - Self-hosted link catcher: drop links into LINE, Slack, Discord or Telegram, send them with a Chrome button, or hand them to an agent; driftnet follows each one to the actual tool or repo, writes a short explanation in your language and puts it on a searchable page that agents can also read over MCP. Jev judges each item's category, kind, worth and whether it looks like prompt injection. Project guide.
• DriveJev - Open System I autonomous-driving decision model (DriveJev-4B) that scores behaviours with probabilities in one forward pass, with JevPilot closed-loop harness. Project guide.
• Drone Autonomy: AI vs Rules - Local benchmark that flies a simulated Crazyflie drone through adversarial events and measures collision rate against decision latency for Jev, other decision models, chat LLMs, rule-based pilots and a perfect oracle. Project guide.
• drouter - Experimental Rust decision API router that sends requests to Jev or an OpenAI-compatible teacher, caches verified answers and promotes audited shadow students. Project guide.
• DSH Decision Router - DeepSeek Harness plugin (Chinese docs) for automatic model selection: when "Auto" is chosen, a separate decision model reads each turn and picks one of 2–6 user-defined routes (provider, model, effort); decision models are local Laya, d1-3B / d1-omni-600M, or — after explicit consent — the System One cloud API (Jev), with an explicit fallback route on timeout or low confidence. Project guide.
• dsh-agent-dispatch (DeepSeek Harness: Jev advises whether to delegate a turn and to which agent) - DeepSeek Harness plugin that asks Jev once before each turn whether the work should be delegated and to which registered agent, and injects advice only when calibrated gates (dispatch band, confidence and margin, risk veto) pass; with no key or a failed call it stays silent; the key is read from the keychain first. Project guide.
• dsh-agent-kit (DeepSeek Harness services for always-on agent workers, with a Jev check service) - DeepSeek Harness plugin kit for always-on agent workers: WebSocket intake, DingTalk, Feishu and WeCom messaging, Claude Code and Codex task delegation, and a ctx.jev service that checks agent results with typed Jev Choice, Noul and Score answers and escalates to a person when a check fails; Chinese documentation. Project guide.
• dsh-auto-review-jev (sperictao) - DeepSeek Harness Auto-permission plugin: TypeSafe Jev reviews each tool call (distinct from xlennart/dsh-auto-review-jev). Project guide.
• dsh-backdoor-guard - DeepSeek Harness guard where TypeSafe Jev (or self-hosted d1) judges every model-requested tool call. Project guide.
• dsh-decision (DeepSeek Harness plugin that gives the chat model a decision_ask tool backed by Jev or a local Clef) - Small DeepSeek Harness plugin that exposes a System One decision model to the chat model as a decision_ask tool taking a state and noul, choice or score questions. The provider is hot-reloadable configuration: hosted Jev (jev-latest at api.typesafe.ai) or a LAN-hosted Clef with the same request shape, switchable per call. The README notes that Jev sends the state to TypeSafe, so egress needs owner sign-off. Project guide.
• dsh-intent-recognition (DeepSeek Harness plugin with a five-layer intent pipeline where Jev scores the candidates) - DeepSeek Harness plugin for intent recognition: keyword rules first, pgvector retrieval to narrow the intent library to the top 5 to 10, Jev probability scoring (high confidence routes straight away, middle goes to a fine-tuned model, low to an LLM fallback), with a scenario and intent management UI, PostgreSQL storage and an API for other services. It can also generate the Jev question config for a scenario. English and Chinese docs. Project guide.
• dsh-jev - DeepSeek Harness plugin registering jev_ask for TypeSafe Jev noul/choice/score (pin GitHub commit; npm name dsh-jev collides with another package). Project guide.
• dsh-jev (iptton-ai) - DeepSeek Harness plugin registering jev_evaluate for mixed TypeSafe Jev questions (distinct from noetion/dsh-jev). Project guide.
• dsh-jev by Bais1031 (DeepSeek Harness plugin that adds Jev decision briefs at critical analysis steps) - DeepSeek Harness (DSH) plugin that calls Jev at critical analysis nodes chosen by deterministic rules and records short probabilistic decision briefs. Shadow mode only records them for later evaluation of decision accuracy and wasted exploration; advisory mode adds briefs that pass protocol and confidence checks to the current request. Neither mode changes the provider, model or reasoning effort. Project guide.
• dsh-jev-context-gate - DeepSeek Harness plugin: data-driven preflight and evidence-aware review/skill gates with TypeSafe Jev (distinct from other dsh-jev* plugins). Project guide.
• dsh-jev-decide - DeepSeek Harness plugin registering jev_decide for TypeSafe Jev noul/choice/score (distinct from dsh-jev / dsh-jev-verify / dsh-jev-prune). Project guide.
• dsh-jev-gate - DeepSeek Harness (DSH) plugin for AgentTeams (Chinese docs) that checks each round's promised deliverables against a frozen baseline and evidence ledger before the team advances, with an optional native-jev decider (TypeSafe Jev) as an advisory or enforcing second opinion. Project guide.
• dsh-jev-goal-judge - DeepSeek Harness add-on: TypeSafe Jev judges goal progress between rounds with budgeted nudges (shadow mode default). Project guide.
• dsh-jev-interceptor - DeepSeek Harness plugin: TypeSafe Jev risk-gates tool calls (shadow→enforce) and optional session-reference scoring. Project guide.
• dsh-jev-kit - DeepSeek Harness plugin with ~23 named TypeSafe Jev advisory judgments (privacy/scope/memory/triage) on budgeted transport. Project guide.
• dsh-jev-loop - TypeSafe Jev judgments at DeepSeek Harness agent-loop gates (no UI dependency). Project guide.
• dsh-jev-mcp (Jev decision tools for DeepSeek Harness: dependency-free client and stdio MCP server) - Dependency-free JavaScript client and stdio MCP server that give DeepSeek Harness (or any MCP client) four Jev decision tools: jev_check, jev_route, jev_score and jev_evaluate, with credential redaction and probability thresholds. Demos cover realistic coding-agent judgments. Project guide.
• dsh-jev-mcp by NingYuanLee (DeepSeek Harness plugin that runs jkudish/jev-mcp and registers its Jev tools) - DeepSeek Harness plugin that reads a key from the DSH credential store, starts the pinned @jkudish/jev-mcp server on stdio and registers its judgment tools (verify, screen, find, rerank and more) plus a jev skill, so the agent can call Jev without a separate MCP config. Providers: TypeSafe, OpenRouter, Cloudflare, Vercel AI Gateway or a compatible endpoint. Project guide.
• dsh-jev-plugin (jackie-cqz) - DeepSeek Harness plugin for TypeSafe Jev noul/choice/score with configurable guardrails and Web UI result cards (distinct from other dsh-jev* plugins). Project guide.
• dsh-jev-plugin (luobosibing2) - DeepSeek Harness plugin wiring TypeSafe Jev into skill/file selection, supervision, corrections, and workspace approvals (distinct from other dsh-jev* plugins). Project guide.
• dsh-jev-prune - DeepSeek Harness plugin: TypeSafe Jev keep/drop pruning plus deterministic receipt compaction (distinct from dsh-jev / dsh-jev-verify). Project guide.
• dsh-jev-thinking - DeepSeek Harness plugin: TypeSafe Jev Choice sets reasoningEffort from the route's offered levels (one question per prompt; MIT). Project guide.
• dsh-jev-verify - DeepSeek Harness plugin with TypeSafe Jev jev_decision plus live jev_verify benchmark (no mock fallback; distinct from dsh-jev). Project guide.
• dsh-jev-zxh - DeepSeek Harness plugin that exposes Jev as a model-callable jev_decide tool for typed decisions. Project guide.
• dsh-judgment - DeepSeek Harness plugin that exposes one structured-judgment service: one state plus a batch of named choice, noul or score questions in, a batch of calibrated probability answers out, with validation, credentials, audit and metering. Project guide.
• dsh-memory-hybrid (long-term memory plugin for DeepSeek Harness with forgetting curves and optional Jev grading) - Long-term memory plugin for DeepSeek Harness that treats memories as having weight and cost: an authoritative key-value layer, a derived vector index, Ebbinghaus-style decay with revival, and mood modulation, on an embedded SQLite engine with no external services. Extraction and grading are rule-based by default; filling in a Jev endpoint and key lets Jev take over the grading choices. Docs in Chinese. Project guide.
• dsh-opencli - DeepSeek Harness (dsh) plugin that drives your own logged-in Chrome or Edge through OpenCLI: 180+ site commands, generic browser primitives, approval gates on writes, watch jobs, and a System One decision layer (Jev or local Laya) for fast checks. Project guide.
• dsh-plugin-jev - DeepSeek Harness (DSH) Web GUI plugin that gives the model a jev_ask tool for calibrated TypeSafe Jev judgments, exposes the same call to sibling plugins as a host service, adds an optional Jev command guard before tool execution, and shows per-chat usage in the browser. Project guide.
• dsh-plugin-jev-compaction - DeepSeek Harness plugin: TypeSafe Jev scores messages for relevance-based compaction (pins constraints/tracebacks; falls back to age-based on failure; MIT). Project guide.
• dsh-session-prefix (DeepSeek Harness session status emojis classified by Jev) - DeepSeek Harness plugin that keeps the session sidebar readable by putting status emoji prefixes on session titles: Jev classifies each session's phase from the latest instruction and output summary (for example debugging, ready for review or done), and the agent can also set badges through a session_prefix tool. Project guide.
• dsh-set-model - DeepSeek Harness plugin where TypeSafe Jev auto-switches model presets and thinking depth each turn. Project guide.
• dsh-subagent-custom - DeepSeek Harness plugin for subagent model control with Jev-based intelligent dispatch. Project guide.
• dsh-system1 - DeepSeek Harness (dsh) package that gives agents a system1_decide tool and plugins a ctx.system1.decide() API for yes/no, choice, score and image-box judgments, with TypeSafe Jev, Laya and Cloudflare Clef providers. Project guide.
• dsh-work-planner (GitHub issue work packages for DSH) - DeepSeek Harness plugin (Japanese docs) that groups open GitHub issues into ordered work packages for the agent's next session, using a decision model such as Jev for size and priority estimates and deterministic rules for dependency order, with no chat LLM. Project guide.
• duckdb-jev - Native DuckDB C++ extension for TypeSafe Jev semantic predicates, Choice classification, and Score rubrics in SQL. Project guide.
• DuoMind - OpenAI-compatible local LLM server that pairs a small local model (System 2) with TypeSafe Jev (System 1) for faster, more accurate coding-assistant answers. Project guide.
• e-jev (self-hosted Jev-compatible API with a pluggable fleet of decision backends) - Self-hosted service that speaks TypeSafe's System One API and serves typed, calibrated decisions from Cloudflare's Clef-Flash by default, with optional backends behind the same wire contract: a vLLM logprob reader, Laya or any other /v1/systemone server. The author reports WorkflowEvals, Banking77 and Spanish-intent results for the default model. Project guide.
• EDDS (event-driven decision system: listeners described in plain language, scored in parallel by a decision model like Jev) - TypeScript library where listeners describe in plain language when they should fire; for each event a decision model scores every listener in parallel and returns one probability per listener, and listeners above their threshold fire. Includes tracing, fallbacks, routing images to a vision model, and a review tool that proposes cheap keyword rules from traces, optionally using Jev to drop coincidental keywords. Project guide.
• Edward - Supervisor for unattended coding agents (CI, overnight runs) that wraps any subprocess with deterministic rules and an optional advisory scorer; EDWARD_SCORER_BACKEND=jev batches all probes into one TypeSafe Jev call per check. Project guide.
• effortless - Claude Code mod that picks the reasoning effort for every prompt with a judge — Jev (TypeSafe key, about 0.25 s), Haiku on your Claude login, or your own OpenAI-compatible model — and shows a prompt-cache countdown with one-click Compact. Project guide.
• effortless-hermes - Hermes Agent port of effortless: a fast judge (Jev/TypeSafe, local model or custom) picks reasoning effort per turn by rewriting existing effort fields. Project guide.
• ego-decision-layer - Pluggable ego lite browser decision layer: one TypeSafe Jev System One call per step with fail-closed guards. Project guide.
• ego-jev - macOS CLI that lets TypeSafe Jev pick each click and field in ego lite browser TaskSpaces from caller-supplied values, with requirement checks. Project guide.
• EHDS Health Data Access Body prototype (rules decide health-data requests, Jev can reject or flag but never approve) - Hackathon prototype (AIM Healthcare Hackathon Vienna 2026, European Health Data Space track) for deciding health-data access requests: fixed rules run the decision order and an Article 54 lexicon, a decision model reads only the request text and can reject a prohibited purpose or flag for review but never approve, and an officer signs off every case. Hosted Jev or a local Kev-4B model can fill the model role. Synthetic data only; the README reports 10/10 correct decisions in the demo set and documents holdout failures. Project guide.
• ElBruno.AI.Decisions - .NET 10 decision SDK with one API for Choose, Score and Assess across Microsoft-Decision-1 on Foundry, local Ollama models and a native Jev package, plus Microsoft.Extensions.AI routing. Project guide.
• English Net (prompt-injection detector of yes/no questions) - Prompt-injection detector whose seven neurons are English yes/no questions answered by Jev in one call; on a locked confirmation set it ties a single good Jev question on ranking, while every verdict names the questions behind it and a blind spot can be patched by adding one sentence. Project guide.
• Engraphis - Local-first, inspectable memory for coding agents (MCP, code-aware recall, WebUI) with optional advisory Jev decisions via BYOK or managed hosted plans. Project guide.
• enowx - Early-release terminal coding agent (one Rust binary, specialist agents, any model provider) with an off-by-default decision model that can use TypeSafe Jev to decide brainstorm-vs-build, which specialist to delegate to, and when to ask the user. Project guide.
• estate-guardian - Overnight AI-agent fleet watchdog where TypeSafe Jev classifies the free-text failures rules miss. Project guide.
• EuLLM - Sovereign, local-first LLM engine (AGPL-3.0, Ollama/OpenAI-compatible APIs) whose v0.7.20 adds a /v1/systemone decisions endpoint that reads option probabilities straight from local models — e.g. its Jev-Style 2B — with an audit trail; no hosted Jev involved. Project guide.
• Eutrya - CLI-first agent (public alpha) with TypeSafe Jev in the decision loop for attention, candidate rubrics, and research micro-steps; offline eutrya demo. Project guide.
• Evaluating and Benchmarking the System One Model Jev - Code for the arXiv paper by Deußer, Sparrenberg and Sifa (arXiv:2609.37647, 2026) benchmarking TypeSafe Jev (pinned jev-1.13.0) zero-shot on 37 public datasets against two open-weight LLMs scored on identical requests; raw responses are on Zenodo so every number can be recomputed offline. Project guide.
• evidence-referee - Claude Code plugin that checks "done" claims and small judgment calls with TypeSafe Jev and keeps receipts (distinct from claude-referee). Project guide.
• EvolveRoute - Local-first Rust LLM routing gateway that picks a model per request with Jev plus a heuristic fallback and logs every decision. Project guide.
• ex_typesafe_ai - Unofficial Elixir client for the TypeSafe AI API — typed structs, Req-based HTTP with retries, and ergonomic noul/choice/score questions. Project guide.
• ExcelPilot - Live Excel agent with Qwen planning and TypeSafe Jev for intent routing, tool gating, and claim checks. Project guide.
• exo-compact (fast Claude Code compaction, with an optional Jev rule check) - Claude Code plugin (beta) that compacts in about a second instead of a minute, quotes your rules word for word and keeps the whole session searchable on disk; an optional Jev rule check, off by default, asks TypeSafe's Jev which of your task's rules each reply breaks. Project guide.
• factory (Claude Code plugin that works a GitHub Project board, with optional Jev decision points) - Claude Code plugin that takes cards from a GitHub Project board through Build, QA and Review lanes, each run by its own subagent (or Codex), and opens merged PRs with evidence; optional Jev decision points start in shadow mode, logging Jev's answer next to Claude's so you can check agreement and spend before letting Jev decide. Project guide.
• fast-browser - Playwright browser automation for Codex/MCP with TypeSafe Jev (or Laya) decisions. Project guide.
• fast-compact - /fc for Claude Code: TypeSafe Jev decides which old tool outputs stay whole; the rest are trimmed with cuts saved to disk. Project guide.
• fast-jev-compaction - Selects tool-call/result pairs to keep, truncate, or remove while preserving retained conversation content verbatim. Project guide.
• fast-jev-compaction (Ayush) - Claude Code plugin that replaces compaction summaries with Jev decisions over tool calls and results. Project guide.
• fast-jev-compaction (Lexus2016 maintained fork: Claude Code compaction by Jev that keeps every call) - Maintained fork of the listed fast-jev-compaction Claude Code plugin: Jev scores each tool call, stale outputs become a one-line note while every call stays, kept content stays verbatim, and community pull requests from upstream are audited and ported. Project guide.
• fast-jev-compaction-opencode - opencode verbatim session compaction via probabilistic keep/drop judgments (local LM Studio judge by default; based on fast-jev-compaction). Project guide.
• fast-jev-opencode - OpenCode V2 plugin: TypeSafe Jev prunes stale tool calls/results on the outgoing request (fail-open; leaves history//compact alone). Project guide.
• fast-jev-opencode (roshan-shaik-ml) - OpenCode v1/v2 plugin: TypeSafe Jev prunes stale/bulky tool calls on the outgoing request (verbatim; distinct from nrdz-labs/fast-jev-opencode). Project guide.
• fast-resume-scoring - Score resume-to-job fit and requirement coverage with one TypeSafe Jev request in under a second. Project guide.
• FastBrain proactive replies for AstrBot (FastBrain-9B decides whether the bot should chime in to a group chat) - AstrBot plugin: a FastBrain-9B System One model reads recent group messages and decides whether the bot should join the conversation unprompted; the session's own chat model writes the reply. Ships in observe mode with an empty group whitelist, so it never speaks until you configure it; rate limits, cooldowns and quiet failure on errors are built in. README in Chinese. Project guide.
• FastBrain-9B (open non-autoregressive System One decision model that answers Jev-shaped requests in one pass) - Open decision model built on a hybrid Qwen3.5-9B backbone with a 116M-parameter joint schema head: give it a JSON state and choice, noul or score questions and it scores every option of every question in a single prefill instead of generating text. Inference, training and evaluation code here; weights on Hugging Face. Not TypeSafe's Jev. Project guide.
• fastbrowse (open-source browser agent where Jev picks each control from the ones the page actually has) - Python browser agent with a Rust core: each page is indexed into the controls it really has and Jev, as a choice model, picks the next one, so the agent cannot click something that is not there. Answers cite verbatim page quotes, and submit, pay, delete and send actions wait for code-level approval with step, time and dollar limits. Project guide.
• fastcampus-jev - Korean Fastcampus Jev tutorial: 13 notebooks + React/LangGraph demo (tool Choice, guardrails, risk gates, RAG) via OpenRouter (MIT). Project guide.
• FastGate - Multilingual EN/UZ/RU RAG helpdesk where TypeSafe Jev is the System One decision layer before the LLM writes, with an independent benchmark. Project guide.
• Fastpunkt - Evidence-first research agent for Norwegian companies (built for the Builderr Signalpost challenge): given organisation numbers it returns one profile per company where every published fact links to its source with retrieval time; official registers anchor identity, and Jev (jev-1.13.0) is used only to select among candidates the code found (links, sentences, job titles) or to give a calibrated same-entity probability — no model-written facts. Project guide.
• FAVA Trails - Git-native, versioned memory for AI agents via MCP (Machine Wisdom AI): drafts are isolated, and since 0.8 the default Trust Gate asks TypeSafe Jev through OpenRouter one typed Noul question with an explicit threshold before a memory is promoted to shared truth. Project guide.
• Feedback Graph (local memory of corrections for coding agents, optional Jev classifier) - Local Docker app that reads Claude Code and Codex sessions, finds the moments you pushed back, and turns them into confirmed lessons your agents get through MCP; Jev is one of the selectable models for classification jobs. Project guide.
• feelings - BAML .feels() / .how() / .matches<T>() typed AI-if methods powered by TypeSafe Jev; upstream licensing unspecified. Project guide.
• Figma Jev Review - Figma plugin that reviews selected interface copy against declared criteria and cites the exact weak TextNodes, so clicking a finding selects and zooms to the original layer; it never writes copy. Project guide.
• finance-decision-evals - Independent evaluation of Jev on ASC 606 revenue-recognition review decisions using synthetic contract clauses. Project guide.
• Finclaude (personal investment-research agent in Claude Code, with Jev probabilities logged and scored) - Personal investment-research assistant built on Claude Code: Claude fetches official sources (SEC EDGAR, Fed, BLS, BEA, investor-relations pages) and computes with Python, while probabilities come only from data and models including Jev; every probability is logged before the outcome and scored by Brier score against a coin-flip baseline. Weekly News, Decision and Lesson pages; read-only broker access, no buy or sell orders. Educational, not investment advice. Chinese docs with English summary. Project guide.
• firefox-jev-mcp - Claude plans; TypeSafe Jev picks Firefox clicks via MCP + WebExtension bridge. Project guide.
• Flick (flick-computer-use) - Local MCP computer-use: agents hand Flick goals; TypeSafe Jev runs browser/macOS observe→decide→act. Project guide.
• FlightBench (fixed-wing landing simulator and benchmark: can you land a plane better than Jev?) - Landing simulator and benchmark on JSBSim physics with a Three.js view: fly the approach yourself, then compare with Jev, LLMs such as Codex, and classical controllers on the same setup. Jev picks control actions from the flight state each step; recordings and published runs include jev-1.13.0 and jev-1.14.0. Project guide.
• Flink Jev (Jev yes/no decisions in Flink SQL with ML_PREDICT) - Community model provider for Apache Flink 2.3 ML_PREDICT: asks Jev one yes/no question per row and returns a probability, a route (yes, no, review or failure) and a SHA-256 digest of the text, with uncertain answers kept apart for review. Project guide.
• FlipJev (black-box red teaming that finds universal adversarial suffixes flipping Jev's decisions) - Code, suffixes and results for a Johns Hopkins and UW-Madison preprint: an evolutionary black-box attack that uses only returned probabilities to find short suffixes which, appended to untrusted content, flip a typed decision model's answer. The authors report that one suffix found on one dataset flips 90-100% of documents in five other tasks. Every table and figure regenerates from released results without API calls. Project guide.
• flowtape - Records a user journey in a running Vite + React or Next.js app and exports it as JSON and a Claude Code prompt to replay and explore; Jev (TypeSafe) finds replay targets visually from screenshots. Project guide.
• Flutter × Jev (measured experiments in driving a Flutter app with Jev) - Measured experiments in driving a real 13-screen Flutter app through marionette: Jev picks every action from options built from the screen, a cheap LLM writes only the text a field needs and referees ties on undoable steps, and every run is judged by what the app stored; includes a GenUI benchmark and UI-churn tests. Project guide.
• fmev (System One API on Apple Foundation Models) - Local server that implements TypeSafe's System One API (POST /v1/systemone, GET /v1/models) on the Apple Foundation Models model built into macOS, so existing Jev clients and SDKs work offline once you change the base URL. Project guide.
• fndds-matcher-jev - Match food descriptions to USDA FNDDS codes via hybrid retrieval plus TypeSafe Jev. Project guide.
• Footwork - Dual-process browser agent: TypeSafe Jev as System 1 in front of browser-use System 2, with a code-owned arbiter and evidence verification (jevdual). Project guide.
• Foreman - Supervises Codex workers with Jev assessments and deterministic steering, retry, and verification policy, with a synthetic demo. Project guide.
• forge (continuous engineering loop) - Claude Code plugin that applies the autoresearch ratchet (one change, one commit, one measurement, keep or reset) inside a team-style process — spec, mapped scope, clarifying questions, tests beyond the happy path and human sign-off — with small tasks run through a test-driven harness and large ones driven to a measurable goal by an orchestrator, parallel subagents and a deterministic judge; Claude does the work while Jev makes cheap judgments such as triage, model routing and failure labelling. Project guide.
• forged - Terminal tool that screens a PDF or image for signs of document fraud: it reads the file's metadata, renders pages to images, and asks Jev through Vercel AI Gateway for a fraud probability and the likelihood of each fixed fraud signal. Project guide.
• Formanator - CLI/MCP for Forma benefit claims; optional TypeSafe Jev chooses benefit/category when you supply claim fields (receipt inference stays on LLM). Project guide.
• Founder CRM - Local SQLite CRM for an early startup's first customers, conversations and promises, with a dashboard, Claude over MCP to propose updates, and optional Jev checks on whether the evidence supports each proposed change before you approve it. Project guide.
• Fovea-4B (open image decision model served over the System One API) - Fovea-4B is a 4B decision model for images: given a photo, document, receipt, chart or screenshot plus a JSON state and a typed question, it returns a probability for every allowed answer in one forward pass. It is served over TypeSafe's System One API shape, so existing System One clients can call it unchanged; the README compares it with Imajev-4B, H2O-Lightning-4B and Jev-Vision on ImageJevBench. Project guide.
• fr (Arindam200) - Ask a codebase question and get relevant files via live parallel TypeSafe Jev relevance. Project guide.
• Frame Engineering for Jev - Agent skill and scripts for designing and evaluating Jev decision frames: evidence in state, atomic questions, answer boundaries and application policy, with ablation, sweep and token-probe scripts. Project guide.
• Frost (model router built on Jev) - Go CLI from Haplab that takes a prompt or markdown task, has Jev assess it, then picks a model, harness and effort level locally from your own catalog, with explanations, alternatives and JSON output; decisions can be recorded and replayed offline. Project guide.
• Fulcrum (.NET library that turns typed questions and ordered rules into one committed outcome for System One models) - .NET library: you declare typed questions and ordered rules, Fulcrum asks a System One model, and one committed outcome (reply, page someone, send to a human) is dispatched to a handler that never sees a probability. Backends cover Jev on OpenRouter or any OpenAI-compatible endpoint, the native Jev API (experimental), and a fake model for tests. Project guide.
• fusion-jev - Local coding evidence plus optional guarded TypeSafe Jev choices for MCP hosts. Project guide.
• FuZev - Jev-compatible Choice, yes/no, and Score decisions alongside text generation on vLLM/ExLlamaV3 from one loaded model. Project guide.
• fuzzy-jev (Rust/Python client for Jev's typed answers with fuzzy rules over the probabilities) - Rust crate, CLI and Python package that ask Jev typed Choice, Score and Noul questions about text and images through OpenRouter's decisions endpoint, then apply a rules file that reads each probability as a fuzzy truth value (AND = min, OR = max) and can draw the rule graph. Includes an Agent Skill that teaches coding agents when to ask Jev. Project guide.
• FYIsoft.Extensions.AI.Jev - .NET Microsoft.Extensions.AI provider SDKs from FYIsoft, including a typed JevClient port of @typesafe-ai/sdk v0.6.0 and an IChatClient adapter for structured judgments, with a support-desk sample where Jev triages tickets and reviews Claude drafts. Project guide.
• Gallant (decision-model linter for code and docs) - Single-binary Go linter that checks code, documents and other text files against fuzzy rules you write (for example, comments that repeat the code), answered by decision models: Jev through OpenCode Zen by default, or Cloudflare Clef and others. Project guide.
• game-rsi - Real-game teacher correction and gated model evolution experiments with PlayJev configuration packs. Project guide.
• gated-llm-pipeline (social-post triage where code owns the flow, Jev gates each step, and an LLM sees only unsure items) - Eight-step triage pipeline for inbound social posts: code owns the control flow, a cheap classifier (Jev, or a deterministic mock) makes the narrow judgment at every gate, and a reasoning model only sees items the classifier is unsure about. Answers are cached, and backtest and QA-eval scripts check the gates. Project guide.
• Gatekeeper - YAML rules plus TypeSafe Jev Choice route Claude Code prompts to the right skill/agent (allow/route/block bands). Project guide.
• Gatekeeper Editorial 30X - Hackathon web app (in Spanish) where Jev answers typed questions about each content piece and code routes it to auto-approve, LLM rewrite, human review or reject; a second flow sorts Luma event sign-ups into programs and decides whether to call them now. Project guide.
• Gavel - Salesforce packages for governed Choice/Score/Noul judgments with TypeSafe Jev, org policy, and an immutable ledger (v1 pre-release). Project guide.
• gbrain-evals - Benchmark suite for the gbrain agent-memory system, including a matched-pair System One report on where TypeSafe Jev (jev-1.13.0) helps or hurts nine memory decision slots. Project guide.
• Gemma 4 Typed Decisions (independent Jev-compatible decision API on Gemma 4, benchmarked against Jev) - Independent implementation that makes Google's Gemma 4 return typed decisions with candidate probabilities (Noul, Choice, Score) through a /v1/systemone API, without generating text, plus a harness that compares it with the official Jev on the same inputs. The author reports 78.0% vs 79.9% accuracy on 3,300 test items with lower latency on an NVIDIA GB10. Japanese-language docs. Project guide.
• genai (maruel) - High-performance multi-provider Go AI package whose Provider.SystemOne() asks typed decision questions across TypeSafe Jev, Cloudflare CLEF/CLEF-Flash, Ollama and local llama.cpp decision models with shared genai.Questions / genai.Answers maps. Project guide.
• genigrep - Ask a codebase question and get only answering source; TypeSafe Jev ranks candidates for coding agents. Project guide.
• Genkit flows on Kubernetes routed by Jev - Go example where each step of a support-ticket pipeline is a Genkit flow running as a Kubernetes pod through senro; Jev decides department, urgency and escalation through Genkit's System One plugin and then reviews Gemini's draft reply. Project guide.
• Genkit software factory gated by Jev - Go example of a software factory on Kubernetes: Gemini designs, plans and writes a Go module as Genkit flows run through senro, and Jev holds every gate, deciding whether the design is ready and whether to ship, rework or hand a human a brief. Project guide.
• ghostel-ai (natural-language commands for Emacs ghostel with a Jev gate) - Emacs minor mode for ghostel's line mode: Jev decides whether what you typed is a shell command or a plain-words instruction and flags destructive input for confirmation; gptel turns instructions into commands for review. Project guide.
• ghtriage - Classifies GitHub issues with TypeSafe Jev typed labels and calibrated confidence; writes are opt-in (ghtriage). Project guide.
• ghwatch - Ruby orchestrator that runs coding, review and finalizer agents off GitHub issues and pull requests, with optional TypeSafe Jev issue screening before work starts. Project guide.
• git-jev-stage - Classifies Git hunks against a one-sentence staging intent with TypeSafe Jev, then stages confirmed blocks. Project guide.
• gitpp-gate - Confidence-gated nudging for the git.pp fold: asks TypeSafe Jev before nudging, fails closed, and logs the decision so automation cannot push without a typed yes. Project guide.
• glimpse - Computer-use harness for the Linux GNOME desktop: a planner model writes the plan, and Jev finds each control, checks each result and decides when the task is done, across LibreOffice, Files, the browser and other apps. Project guide.
• gloss (streaming, annotated review tours for pull requests and local diffs, with Jev judging each file) - Turns a pull request or local diff into a guided review tour in the browser: ordered groups with summaries and notes pinned to the lines that matter, streamed stage by stage. An LLM defines groups and writes notes; Jev makes the per-file calls (which group, generated, tests-only, and whether a file touches auth, stored data, secrets or personal data). Requests are cached locally. Project guide.
• gm-autoplayer - Agent that plays browser games such as slither.io from screenshots through the Playwright MCP server; pick an LLM brain, a Jev brain that answers quick typed questions about each frame (danger, nearest food) for game rules to turn into moves, or a hybrid with an LLM strategist. Project guide.
• gmail-auto-cleanup - Python tool that sorts a Gmail inbox with a System One model — TypeSafe Jev (tested on ~27,000 emails) or experimental local Laya — answering typed yes/no questions per email while plain Python rules act; actions are archive + label with a 7-day hold before Trash, and the OAuth scope cannot permanently delete mail. Project guide.
• go-decide - Go CLI that puts a choice or ordered rubric to a System One model — TypeSafe Jev by default, Cloudflare Clef opt-in — and prints a JSON outcome whose exit code tells your script to act, ask a person or escalate; it gates and labels decisions but never approves them. Early development, no tagged release yet. Project guide.
• go-discern - Small typed Go decision layer for the TypeSafe Jev and Cloudflare Clef System One APIs, with validated observations and zero third-party runtime dependencies. Project guide.
• go-jev - Go SDK and jev-cli for TypeSafe Jev typed Noul/Choice/Score (explicit key; distinct from jevgo and TypeSafe Go). Project guide.
• Godot Jev (Godot 4 addon wrapping the Jev System One API) - Godot 4.7+ addon that exposes Jev as an autoload: choice, score, noul and box questions about game state, batched in one request, with native Godot inputs such as Dictionary, Image and Texture2D. Project guide.
• godot-jev-actions (Jev action nodes for Godot 4) - Godot 4 addon with designer-facing Jev nodes: an ActionCatalog allowlist, an ActionPicker that turns a sentence or game state into one allowed action with arguments, and a FreeTextChoice that maps typed text to an authored choice. Nothing is generated. Project guide.
• GoEventBus - High-performance Go event bus with deterministic rules, a decision cache, and an optional JevSelector that asks Jev (via OpenRouter's Decisions API) to choose an event type only on a cache miss. Project guide.
• goinfer - Pure-Go, no-cgo local LLM inference (library and server) whose POST /v1/systemone answers TypeSafe's decisions wire shape so clients such as jevx work against it; it scores option labels with the served model or loads a trained decision head — TypeSafe's hosted model is not included. Project guide.
• gojev - Go harness for System One decision models — TypeSafe's hosted Jev, Jev via Vercel AI Gateway, Cloudflare Clef and in-process open-weight Kev — with typed Choice/Noul/Score question builders, calibrated answers, gating helpers and a CLI. Project guide.
• Gomoku × Jev - Five-in-a-row against Jev on a Three.js 3D board (README in Chinese): local heuristics propose up to ten tactically labelled moves and Jev picks one with a Choice question, showing the pick and its confidence in the status bar. Project guide.
• Google ADK TypesafeClassifier (route ADK agent workflows with Jev) - Built-in integration in Google's Agent Development Kit for Python (v2.13.0+): TypesafeClassifier answers Choice, Binary and Score questions with TypeSafe's Jev, and the experimental ClassifierNode routes a workflow on the answer, with route labels checked against the graph when the workflow is built. Project guide.
• gpui-agent - Jev-driven native GPUI testing over accessibility trees (experimental, macOS-first, instrumented apps). Project guide.
• gr-jev (paper code: asymmetric probability allocation when several choices are correct) - Code and data for the arXiv paper "When Multiple Choices Are Correct: Asymmetric Probability Allocation in Jev and Other Decision Models": experiments showing that when two options are both correct, or five options are identical, Jev and other decision models split probability unevenly, which breaks threshold-based multi-select. Project guide.
• granite-switch-systemone (vLLM System One endpoint) - vLLM endpoint plugin that serves a Granite Switch checkpoint's decision adapter through both OpenAI's Decisions API and Jev's System One API on the same server and GPU cache as chat completions. Project guide.
• graphify-jev - Codebase knowledge graph with Tree-sitter AST topology plus TypeSafe Jev semantic navigation for coding agents (no vector DB). Project guide.
• Graphlin - Live architecture and activity diagrams for Claude Code or Codex while they explore code, with optional TypeSafe Jev classification. Project guide.
• Greenlight (Pi risky-command checks by Jev, escalated to the orchestrator agent) - Pi extension that builds on the listed pi-jev: Jev checks each risky command in about a third of a second, and a flagged command is paused and sent up the chain to the orchestrating agent (Traycer, a parent Pi or any A2A orchestrator) to approve, deny or ask a human; warn-only by default. Project guide.
• grep2jev - CLI that lets a coding agent hand file triage to Jev: pipe in candidate files, Jev scores each against the request in parallel (Python sent as AST skeletons), and the agent reads only files that pass. Project guide.
• grev - "Thinking coreutils": Unix filters (grev, isv, tagv, rank, sortv, pickv) that ask TypeSafe Jev typed questions, so output is always your input with labels or scores only as explicit columns. Project guide.
• grill-with-jev (agent skill that grills a plan in rounds, with Jev deciding) - Agent skill from 1arley's skills collection: the host agent builds a design tree, asks each round of open questions with a recommended answer, and Jev, acting as the user's automated stand-in, decides through schema-validated answers; ends with a decision record of decisions, provenance, assumptions and open risks. Budget, round and session limits are enforced by the script. Project guide.
• Grok Bot Jev - Reference router and skill that use TypeSafe Jev to choose reuse/stop/cap/ask-human actions before expensive Grok Bot work. Project guide.
• grok-jev-guard - Typed preflight/approval layer for Grok Bot: local policy plus TypeSafe Jev for ambiguity (shadow-first; distinct from Grok Bot Jev). Project guide.
• grokbot-jev-skills - Ten Grok Bot agent skills, adapted from hermes-jev-skills, that hand small decisions — which search results to read, passage filtering and injection screening, mailbox sort, support triage, routine wake-ups, transcript cutting — to TypeSafe Jev through a jev CLI; fail-open, with documented redaction. Project guide.
• Groot 1 (open typed-decision model in the System One format) - Open 149M-parameter typed-decision model that answers System One choice, score and noul questions with full probability distributions over states of up to 64k tokens, with a pure Bend implementation of its forward pass; the author reports it beating Jev on their 2,000-question test split. Project guide.
• guardrails-md - Berget AI plugin for opencode, pi and Claude Code (plus a git pre-commit hook) that scores every bash command with a System One decision model against your repo's guardrails.md before it runs; defaults to Berget's berget/bev model, and any /v1/systemone-compatible endpoint (TypeSafe settings as fallback) can be used. Project guide.
• guesswork - zsh history autosuggestions ranked by TypeSafe Jev instead of prefix match. Project guide.
• Gut - Elixir DSL for typed decisions in ordinary control flow (Gut.feel/3 returns an Elixir value) with a Gut.ReqLLM.Jev adapter for TypeSafe Jev, plus other LLM, Ollama, and custom adapters. Project guide.
• Gut Feel Benchmark - Benchmark of snap probabilities at deadlines from a quarter second to 16 seconds: 703 boolean items in four packs (poker equity, base-rate puzzles, heart disease, Hacker News outcomes), sent to Jev, other decision models and chat models through AI SDK's experimental_decide. Project guide.
• gutfeel - Usability testing for MCP servers: a participant that doesn't deliberate — TypeSafe Jev by default — walks your tools on realistic, hint-free missions, and gutfeel draws the tree of every path (right tool, wrong turn, dead end, error, loop) with the description sentence behind each failure; pre-alpha, with gutfeel live runnable today. Project guide.
• gutsy - Local calibrated gutsy-0.8b decision model with Jev-compatible /v1/systemone and Decisions APIs on CPU (independent of hosted TypeSafe Jev). Project guide.
• HA Jev Autopilot - Home Assistant: TypeSafe Jev decides per-room actions; code acts with ask-first for risky devices (distinct from HA-Jev). Project guide.
• HA_SystemOne - Polish-language Home Assistant command layer: local embedding and reranker models narrow about 100 possible actions to five, Jev picks one or none with a confidence, and the action runs only at 0.8 or above; otherwise it asks back. Project guide.
• hak-jev-plugin - Hedera Agent Kit gate: TypeSafe Jev judges sensitive tool calls (v0 SaucerSwap) before sign; HCS receipts. Project guide.
• Hammertime - Local live Formula 1 strategy desk: a strategy engine owns the numbers (tyre model, pit loss, Monte Carlo), while System One models such as TypeSafe Jev read race-control messages, gauge rain risk, and weigh evidence into box-now or stay-out calls. Project guide.
• Handoff Graph - Issue-routing experiment that stores handoffs in FalkorDB and asks TypeSafe Jev one narrow question per candidate owner. Graph memory cut human escalations from ~two-thirds to under 10% on held-out VS Code issues in the author's study. Project guide.
• harness-router - Fast decision routing for agent harnesses — native MCP with Jev for tool selection and MCTS for multi-step decisions. Project guide.
• HarnessMM (paper code: reuse of agent decisions for harness token efficiency, with Jev runs) - Research code for an arXiv paper on making agent harnesses cheaper by reusing decisions: a slow model writes future tool actions into an executable decision graph, guarded reuse checks each tool result before replaying a stored action, and recursive self-improvement tunes the setup; ALFWorld runs include a Jev decision runtime. Project guide.
• HarnessRouter Community Edition (self-hosted unified API for 30 agent harnesses, including a System One harness for Jev) - Self-hosted, Apache-2.0 gateway that runs Codex, Claude Code, Hermes, Pi, DeepSeek Harness and about 25 other agent harnesses behind one Unified Harness Protocol API, with sessions, streaming, files, cancellation and explicit end states. One of the harness bases is System One, which drives the open System One Harness with Jev as the decision model. Project guide.
• Harness² (recursive agent harnessing) - Python framework that improves an agent's harness (prompts, guardrails, memory, tools, skills, scripts) task by task while the model weights stay fixed, with OpenCode and Codex executors, three professional benchmarks, and optional Jev selection of the best final output when it runs k > 1 candidates. Project guide.
• HearMemory - Shared multi-agent coding memory: TypeSafe Jev checks claims against recorded evidence before the next agent trusts them. Project guide.
• HekaJev - Reproducible Git-history analytics powered by TypeSafe Jev classifications. Project guide.
• Herdr macOS plugins - Four macOS plugins for the Herdr agent terminal: open files from agent output, keep the Mac awake, show CPU, disk and folder size, and switch Jev on or off per Claude Code or Codex session through jev-gateway. Project guide.
• Herdr-Jev - Plugin for the Herdr terminal workspace and AI-Harness that uses TypeSafe Jev System One (~260–340 ms) for multi-model triage — architectural complexity, research need and reasoning effort — then routes turns to Codex, Claude, Kiro and other peers with calibrated latency deadlines, Triad plan/implement/review orchestration, quota awareness and resumable runs. Separate from the listed alexzfe jev-herdr. Project guide.
• here-we-go-jev - Local playground/test bench for TypeSafe Jev: typed questions, calibrated probabilities, side-by-side LLM baseline, and mock backend (Bun). Project guide.
• Hermes Adaptive Effort - Hermes Agent plugin that asks a separate scorer (Jev by default) to judge each task and sets the outgoing request's reasoning effort. Project guide.
• Hermes Bots plugin - Hermes agent plugin behind the Hermes Bots mobile app: native cards and forms, phone alerts, voice, model routing, and a card guard where Jev decides in about 0.2 s whether a prose reply should become a card. Project guide.
• Hermes GPT (Jev decision bridge) - MCP connector that gives ChatGPT controlled access to a Hermes Agent machine, now with an optional Jev bridge: three typed tools for yes/no, choose and score judgments, plus an advisory Jev review score for dispatched work. Project guide.
• Hermes Jev Skills - Agent skills and jev CLI using TypeSafe Jev for model routing, memory filtering, compaction, skill selection, triage, and computer/browser action choice on Hermes, Claude Code, and Codex. Project guide.
• hermes-adaptive-model-router - Shadow-first adaptive model router plugin for Hermes Agent: a redacted, current-turn routing dossier goes to TypeSafe Jev, which chooses between a fast model and a high-capability model; decisions are logged content-free while the gateway's configured model still runs. Project guide.
• hermes-jev (ourines) - Hermes Agent decision sidekick: explicit Jev tools and a bundled skill via TypeSafe, Cloudflare, or OpenRouter backends. Project guide.
• hermes-jev-curator - Hermes plugin: TypeSafe Jev typed skill-relationship judgments and safe curator archive/guard plans (experimental 0.1.0). Project guide.
• hermes-jev-helper - Hermes Agent plugin: TypeSafe Jev intent classifier picks a source-first route (skill/research/…) before the LLM turn. Project guide.
• hermes-jev-lesson-gate - Hermes plugin: TypeSafe Jev judges when memory/skill background review should run instead of a turn counter. Project guide.
• hermes-loopguard - Hermes Agent plugin with cheap typed semantic gates (TypeSafe Jev via typesafe-sdk) plus zero-token change gates and computer-use gates to stop automations from burning tokens in loops. Project guide.
• hermes-swift / Alpine Swift (fast, verified browser tasks for Hermes Agent) - Browser plugin for Nous Research's Hermes Agent on Ubuntu: Jev picks each step from the page's indexed controls, Inception's Mercury 3 writes form values, tasks count as verified only when a code check on the final page passes, logins and cards come from Hermes's vault so no model sees a password, and irreversible steps are reviewed. Project guide.
• Hev - Working Jev-compatible Decisions API answered by human visitors on a live page; swap the Jev base URL to test an integration against human answers (not for production). Project guide.
• himalaya-jev-mail-classify - CLI that classifies Gmail threads via himalaya + TypeSafe Jev (OpenRouter) and applies labels (dry-run default). Project guide.
• Hirelog (shared, evidence-backed hiring memory for job-search agents, with Jev write checks) - MCP server and live service where job-search agents share typed, evidence-backed notes about employers and hiring systems (application forms, location rules, interview rounds); an optional Jev check at write time asks cheap typed questions, such as whether a note names a private person, before it is stored. Project guide.
• HiRoute - Local-first agent routing engine with TypeSafe Jev decision extensions for branch selection. Project guide.
• HN Jev Sorter - Chrome extension that sorts Hacker News comments into custom buckets in real time with Jev probabilities. Project guide.
• hndelegate (Hacker News curator that sends stories matching your Jev decision trees to Telegram) - Go service that polls new Hacker News stories every two minutes, fetches their content, evaluates each against one or more customisable decision trees stored in SQLite with Jev answering each node, and sends stories that match a tree to a Telegram chat or channel. Project guide.
• Hollow Knight + Jev - Source-only research prototype that patches Hollow Knight with a C# bridge exposing bounded numeric observations, lets Jev choose movement and nail strikes as separate questions against Gruz Mother in Godhome, and applies short input leases, with a spending ledger and a local dashboard. Project guide.
• Honeytongue - NPC persuasion engine: TypeSafe Jev judges whether player dialogue convinced a character. Project guide.
• hono-jev-router - Experimental Hono router that matches HTTP requests to plain-English descriptions with TypeSafe Jev Noul judgments. Project guide.
• hookgate - Claude Code/Codex hooks: TypeSafe Jev gates risky Bash and unverified completions (audit/fail-open). Project guide.
• Hopwise (knowledge-graph traversal benchmark) - Benchmark (README in Vietnamese) where Jev picks which edge to follow at each node of a Neo4j knowledge graph built from Wikidata and SEC EDGAR: 26 of 60 questions right (43%) in a 4.2 s median for $0.07, against 25% for Gemma 4 12B and 27% for gpt-oss-20b. Project guide.
• hsom (jevdo) - Strict-allowlist harness that lets Jev pick bash commands, flags and paths from a TOML decision tree. Project guide.
• hunch - Python verbs (classify/score/check/pick/rank/where) over scalars, lists, and pandas columns backed by TypeSafe Jev (hunch-jev). Project guide.
• hunch (Rust CLI for asking Jev yes/no, multiple-choice and scoring questions) - Small Rust CLI that sends a state plus one typed question (yes/no, choice or score) to Jev and prints the answer with probabilities, through TypeSafe's API or OpenRouter. The author describes it as a personal learning project with no maintenance promise. Project guide.
• hunch (Unix filter) - Go command-line filter that sorts, filters and routes text by meaning: pipe in logs, tickets, reviews or diffs, ask a plain-English question, and Jev answers it for every line as yes/no, a category or a score with a confidence your next command can use. Project guide.
• Huncho - Rust server that implements the Jev /v1/systemone API for self-hosted System One models (ONNX, Laya, Kev, Clef runtimes, plus a mock), so the TypeSafe SDK can point at your own server. Project guide.
• hyalo - Rust CLI for Markdown knowledgebases (LLM Wiki pattern, frontmatter queries, links) whose hyalo-tidy agent skill can, when you explicitly ask, have TypeSafe Jev suggest missing types and existing folders for selected documents — audit-only suggestions the agent checks against local schemas. Project guide.
• Hybrid Brain (classroom demo: Jev routes, an LLM answers) - Live classroom demo of System 1 decision models: you type a task and two routers pick which GWDG supercomputer model should answer, Jev (via OpenRouter) or a free local CLM-8B answering typed Choice, Noul and Score questions with probabilities versus an LLM generating its routing decision as JSON, showing decision time, probability bars, cost and whether the JSON was usable; a gatekeeper trusts Jev only above a confidence slider. Project guide.
• hybrid-search-rag (measured RAG ablation over near-duplicate docs, with Jev as a reranker) - Small local RAG pipeline (BM25 plus dense vectors fused with RRF, an LLM reranker, cited answers and a deterministic citation checker) measured stage by stage on a synthetic corpus full of near-duplicate documents; one experiment swaps the slow LLM reranker for Jev, asking one yes/no question per passage, and reports quality, latency and cost with confidence intervals. Project guide.
• ICDLens - Python SDK for ICD-10 lookup where TypeSafe Jev scores catalogue candidates for human-reviewable coding suggestions. Project guide.
• imajev - Open Jev-style multimodal typed-decision models (photos+records+text → options with can't-tell). Project guide.
• infer (self-hosted inference gateway that routes LLM, Decisions and System One / Jev calls through one place) - Rust gateway for watching and managing inference API calls centrally: native passthrough to providers including TypeSafe's /v1/systemone, model resolution for chat, OpenAI Decisions and System One requests, translation between Decisions and System One question types, and support for local Jev-style models on llama.cpp. Configured with a TOML file and separate proxy and control tokens. Project guide.
• ING AI Guild Spring AI demos (compliance RAG where Jev screens and reranks passages) - Two Spring AI demos shown at the ING AI Guild meetup on 27 October 2026, built on use cases ING runs on its GenAI platform: a compliance assistant for new products in a naive RAG and a modular RAG version, and a customer-support agent. In the modular version Jev screens the retrieved passages for injection, contradiction and relevance and reranks them before Gemini answers. Slides included. Project guide.
• InspireJev - Browser-automation execution toolkit (Chinese and English docs) where the main agent plans and verifies while TypeSafe Jev picks Playwright actions from candidates actually observed on the page; GPT, Pi and MCP entries, with optional local /v1/systemone deciders in the 1.1 RC. Project guide.
• IntelligentDocumentProcessing (Jev + Haiku) - Hybrid document processing where TypeSafe Jev decides type and completeness and Claude Haiku extracts, benchmarked on 250 documents. Project guide.
• Intent Race (side-by-side race of Jev against a frontier LLM classifying 1,000 search keywords by intent) - Web demo that classifies 1,000 search keywords into informational, commercial, transactional and navigational intent with Jev (four yes/no questions per keyword) and races it against an LLM, adding up speed and cost live. By default the LLM side is simulated from measured call times and list prices, which the README states plainly. Project guide.
• Intent-Router - Agent skill that compiles vague requests into typed IntentSpec contracts (probe/ask/halt) as input for Jev/Laya routers. Project guide.
• IntentLock - Service that uses TypeSafe Jev (TypeSafeClient, default https://api.typesafe.ai) as the “smart” decision path for intent locking / approval-style checks alongside other LLM lanes. Project guide.
• intune-why (the few fixes that clear most Intune noncompliance, with Jev classifying custom settings) - Reads Microsoft Intune compliance data through read-only Microsoft Graph and builds a fix plan: each step picks the setting whose fix makes the most additional devices compliant, with a what-if view of projected compliance; rules map standard settings, and custom compliance settings that no rule matches can optionally go to Jev for a remediation category. Live demo with a fictional tenant. Project guide.
• invalidate - Semantic TTL for agent memory: TypeSafe Jev checks every stored fact against new evidence; code marks superseded memories without rewriting text. Project guide.
• IronBee Express - Goal-driven browser agent: each step is one TypeSafe Jev decision over the page's offered controls plus one IronBee DevTools call that acts and returns the next snapshot, so model output never becomes a selector, coordinate or script; after the run it reviews the final page and API responses (and backend traces with IronBee), and runs can be recorded and replayed without the engine. Project guide.
• IronBee Gamer - Plays browser games live (2D canvas, Phaser, PixiJS, Cocos, HTML): a perception adapter turns what the page draws into a small JSON state and a decision engine picks every input — hosted TypeSafe Jev with your key, or Laya fine-tuned per game and run locally — with a web UI showing each state, action and probability, a coding-agent trainer for new games, and a shared Hugging Face library. Project guide.
• is-lesion - npm package that evaluates MRI and other medical scan images against Hugging Face dataset labels with Jev 1.14 (Noul, Choice, Score and Box), or local OpenJev vision models, returning a label, probability and bounding box; can score dataset rows directly. Project guide.
• is-malicious - Scans source/config/CI files with TypeSafe Jev for deceptive or data-stealing behavior and points to suspicious lines. Project guide.
• J++ - Experimental language and Rust runtime for composing semantic Jev questions with exact methods as values; offline fixtures included. Project guide.
• J3v - Edge-compiled System One decision runtime ('J3v is to Jev as k3s is to k8s') with Rust compiler/firmware demos. Project guide.
• japanese-jev-lint - Go CLI (jjl) that flags Japanese sentences with TypeSafe Jev Noul probabilities (plus local です/ます regex); no rewrite generation. Project guide.
• Jarvis (Amphion) - Research system and paper ("Coordination Deserves a Brain of Its Own") for full-duplex voice agents: JEV decides each turn (answer, memory, worker, clarify, control, ignore or wait) so you can keep talking while background tasks run. Project guide.
• JarvisCore - Python multi-agent runtime (peer-to-peer mesh, durable state, integration atoms) with a typesafe extra that injects a shared Jev client as self.decisions and can route subagents, task complexity and RAG passages with Jev. Project guide.
• JCR - Jev Capability Resolver: TypeSafe Jev searches a nested capability tree and returns deterministic command documentation for agents (MCP + harnesses). Project guide.
• jear - Rust Jev-routed client for NEAR AI Cloud inference and IronClaw agents using budget/quality/sensitivity decisions. Project guide.
• Jeeves (PostHog) - Open 9B Jev-like reasoning classifier (noul/choice/score) with a Jev-compatible API and training code; independent of TypeSafe-hosted Jev. Project guide.
• jeff (Claude Code router) - Claude Code plugin that uses Jev to route every subtask: model tier, Claude or Codex, which skill, write access, and parallelism, then asks Jev to verify each result and escalates misses. Project guide.
• Jeff (firelex) - Jeff: an open 0.8B "System 1" decision model (MIT code, Apache-2.0 weights on Hugging Face) that uses the same request format as Jev, with swappable LoRA adapters (guard, triage, tools, grounding, spam…), GGUF files for llama.cpp, and Jeff-Code; independent and not affiliated with TypeSafe. Project guide.
• jeff (Viperwow) - Single-binary Rust classifier server with an admin page: one /v1/systemone-shaped API in front of Jev-compatible models — TypeSafe Jev in the cloud, a local Contrastive-LM (CLM) it installs, or Laya/PipeLLM/any compatible server — with access keys and a playground. Project guide.
• Jeffort - Claude Code plugin: TypeSafe Jev scores each prompt and sets Claude effort without touching the model (MIT). Project guide.
• jegrep - Semantic grep CLI: describe code intent and get Jev-ranked matches via TypeSafe or OpenRouter, without an embedding index. Project guide.
• jes - Open-source guardrails for AI agents that ask Jev to judge prompts, retrieved content, tool calls and responses for prompt injection, jailbreaks and secret/PII leaks. Project guide.
• jev (drpaneas) - Go package for calling TypeSafe Jev System One typed questions. Project guide.
• jev (okooo5km) - Stdlib Python CLI + Agent Skill for TypeSafe Jev yes/pick/score via TypeSafe API or OpenRouter (distinct from typesafe-cli / typesafeai-cli). Project guide.
• jev (polidog) - Unofficial Rust CLI for TypeSafe Jev across TypeSafe, Cloudflare, and Vercel providers (CLI shape follows stefafafan/jev)—distinct from feder-cr/jev (jevos). Project guide.
• jev (stefafafan) - Unofficial provider-neutral Go/Unix CLI for TypeSafe Jev (Choice/Noul/Score) via TypeSafe, Cloudflare, or Vercel. Project guide.
• jev (taifoon-io) - Grade an AI agent job with TypeSafe Jev: fact checks in code, four closed questions, receipt with probabilities, optional on-chain records (@taifoon/jev). Project guide.
• Jev + Claude Code case study - Independent learning project testing what "Jev + Claude Code" means with two working programs: Triage Desk, a support-ticket app asking Jev six typed questions per ticket in one request with low-confidence cases going to human review, and Jev Guard, a Claude Code PreToolUse hook asking five redacted noul questions about ambiguous shell commands that never answers allow; both have offline evals and keyword baselines. Project guide.
• Jev + Claude email triage guide (step-by-step: set up Jev with Claude Code and build a read-only Gmail follow-up sorter) - A guide for people who don't write code: install Claude Code and the TypeSafe skill, try Jev on sample emails, connect Gmail read-only, and get a ranked report of the last two weeks sorted into piles such as Reply needed, Waiting on them, Check and Possible phishing. Includes a guided /setup command, customisation and an honest-testing chapter. Tested on Windows; Mac steps written but untested. Project guide.
• Jev 1.14 comparison benchmark (etsabary) - 1,136 deterministic text decisions across 25 families run on eleven services (Jev 1.14 and 1.13, Luna Decisions, Luna, Mercury, Gemini Flash-Lite, Qwen, Granite and others) with matched diagnostics, latency percentiles and versioned results. Project guide.
• Jev Account Filter - Screens video-platform accounts for dataset curation with Jev title ranking followed by multimodal account-category decisions. Project guide.
• Jev Ad Block - Chrome MV3 ad blocker that combines network rules with Jev judgements of candidate page elements. Project guide.
• JEV ADK - Agent Development Kit patterns for TypeSafe Jev: guardrails, dual-brain routing, and PR triage blueprints (System One decisions in code). Project guide.
• Jev Agent Builder - Single-file visual builder for Jev requests: compose Choice, Score and Noul questions with {{parameters}}, validate against documented limits, and export the exact /v1/systemone JSON, cURL, a TypeScript function or a reusable agent definition. Project guide.
• Jev Agent Control - OpenCode V2 plugin that uses Jev to route requests between Plan, Build, and custom primary agents, with automatic handoff at execution boundaries. Project guide.
• Jev Agent Kit (cloud-f1) - Claude Code plugin that prunes long Bash output with local rules plus an optional Jev relevance pass, keeping originals recoverable. Project guide.
• JEV AgentNetwork - Event-driven multi-agent network where TypeSafe Jev reassigns pending work whenever any agent finishes. Project guide.
• Jev Ants (a thousand Jev-steered ants that form the word or picture you give them) - Browser simulation where every ant is its own Jev agent: it sees its nine neighbouring cells, how many ants each still wants and its neighbours' headings, and Jev answers which way to move and whether to settle, so about a thousand decisions turn a random scatter into a typed word or uploaded image; an image overseer and spend and time caps are built in. Project guide.
• Jev Ants by gkvoelkl (Streamlit ant colony where every ant asks a decision model what to do, with a German and English tutorial) - Streamlit game: you are the queen and give the colony an order in plain text, and each ant asks Jev (or a local Ollama model) which next step fits its situation. A step-by-step tutorial in German and English builds the app from one ant upward. Project guide.
• Jev as an LLM Safety Guardrail (USC study: Jev as input filter and selective output reviewer for a small LLM) - Reproducibility code and data for a USC paper testing Jev as a guardrail around a Qwen2.5-1.5B pipeline on 970 AdvBench and XSTest prompts: the LLM alone, a Jev input gate, and a selective setup that blocks clearly harmful requests, passes clearly safe ones and reviews the answers in between. Reports how many harmful prompts got answers and how many safe ones were refused, with threshold sweeps rebuilt from cached Jev scores. Project guide.
• Jev Behavior for Unity - Guarded asynchronous decision nodes for Unity Behavior using a server-side DecisionPacks gateway for Jev-style decisions. Project guide.
• Jev Bench (cocodedk) - Configurable workbench for repeatable TypeSafe Jev latency and accuracy experiments, with published results at jev-bench.cocode.dk. Project guide.
• Jev bias audit (134,772 paired answers testing how one identity word shifts Jev's decisions) - Counterfactual bias audit of jev-1.13.0 and jev-1.14.0: prompts that differ only in one identity sentence (82 identities across nine attributes, 50 names, placebos) over hiring, lending, trust and other templates, with bootstrap intervals, Benjamini-Hochberg correction and a birth-month placebo floor. The author finds explicit labels move decisions by 3 to 15 points while names barely do. Project guide.
• JEV Book Tags - Calibre plugin for AI-assisted book tagging with TypeSafe Jev (single-book and batch; GPL-3.0-or-later). Project guide.
• Jev Browser (openqa-cn) - Indexed Playwright automation: TypeSafe Jev chooses control/op; CodexQA skill with replay and HTML reports. Project guide.
• Jev Browser (tontoko) - Integrates Jev field selection and source-backed extraction with a Playwright SDK, CLI, MCP server, and explicit assertions. Project guide.
• Jev Browser (Ying-Kai-Liao) - Turns caller-defined browser goals into Jev-selected Playwright actions with a library, CLI, and MCP server. Project guide.
• Jev Browser Agent - Bun and TypeScript browser agent with an indexed action space: a typed decision model (Jev, local or cloud) picks the operation and element each step, a chat helper writes a checklist, and a live inspector streams Chromium as it works. Project guide.
• Jev Browser MCP (saurabhgangurde) - Local MCP server that lets a coding agent run bounded browser tasks in a dedicated Chrome window: browser_task follows a goal through Jev-selected page actions drawn only from the request and live page evidence, with the full Chrome DevTools MCP operation set also exposed. Project guide.
• Jev Browser Skill - Claude Code/Codex skill where TypeSafe Jev chooses browser operations and targets from a code-built element table (reference implementation; explainer site included). Project guide.
• Jev Bug Hunter - Bounded first-pass bug hunting for source files powered by TypeSafe Jev — a quick sanity check, not a replacement for tests or review. Project guide.
• Jev by Example - Ten runnable JavaScript lessons on agent decisions between steps; Jev supplies typed judgments and code owns policy (offline fixtures by default). Project guide.
• Jev Chromium - Electron browser agent that collects data across pages: Jev picks each click, focus, navigation or done step from the page's actionable elements, an OpenAI-compatible text model plans, fills and verifies, and results export as a cited table or CSV. Project guide.
• Jev Classificador (Obsidian note classification with Jev) - Obsidian plugin that sends one request to Jev to pick a field (13 options) and subfield (89 options) for the current note from a described taxonomy in English or Brazilian Portuguese, writes them to frontmatter, and can add 1 to 3 topic tags from a closed wordlist. Project guide.
• Jev Classification for n8n - Adds typed classification, scoring, yes/no checks, and batched questions to self-hosted n8n, with configurable review routing. Project guide.
• Jev Classifier for Obsidian - Obsidian plugin: classify note properties with TypeSafe Jev using an editable guide (one note or whole vault). Project guide.
• jev Claude Code plugin (yinjs) - Claude Code plugin that hands small frequent decisions to Jev at six points — permission requests, prompt-injection warnings on fetched content, a 'done' re-check, prompt nudges, and jev-route effort/model routing — each one ~1 s call that changes nothing when Jev is unsure; Jev via OpenRouter (default), TypeSafe or Vercel AI Gateway. Project guide.
• Jev Claude Code Starter Kit (Jev setup, tests and templates for sales systems) - Starter kit for adding Jev to an AI sales or support system with Claude Code or Codex: setup guide and prompts, a tested Python client, a 300-reply test set with scoreboard and cost calculator, ten decision templates, a Claude Code skill and five worked examples, all runnable offline against a fake Jev. Project guide.
• jev CLI (vilinski) - Rust CLI that sends typed questions to Jev and other System One providers and prints JSON answers. Project guide.
• Jev Code Reviewer (egma-ai) - Local PR triage: TypeSafe Jev prioritizes changes; OpenAI explains; Chrome overlay posts nothing to GitHub. Project guide.
• JEV Codex Pilot - Local Codex Kanban workspace with Jev-model routing, contextual compaction, and ticket lifecycle control. Project guide.
• Jev Consensus Gate - Agent action gate that asks Jev about each proposed action three independent ways (safety, legitimacy, adversarial) and only allows it when all three agree; any block wins and any escalation goes to a human. Project guide.
• Jev Content Moderation - Content moderation demo with asymmetric approve / shadow / remove thresholds on Jev probabilities, so irreversible removal needs more certainty than approval. Project guide.
• Jev Cookbook (Datawhale) - Chinese System One/Jev cookbook: notebooks, 18 recipes, docs translation (CC BY-NC-SA 4.0 — free educational / non-commercial; not OSI open source). Project guide.
• Jev custom connector for Power Platform (Power Automate and Copilot Studio) - Open-source Power Platform custom connectors for TypeSafe Jev and Microsoft-Decision-1: define yes/no, choice and score questions in the designer and use the typed answers as dynamic content in Power Automate flows and Copilot Studio agents. Project guide.
• JEV Damocles for AstrBot (group-chat plugin where Jev decides whether the bot should speak) - AstrBot plugin (Chinese) that hands the 'should the bot reply to this group message?' decision to a Jev / System One decision model plus an explicit group-chat state machine: messages are normalized, grouped into topic threads and merge windows, then asked in one multi-question call and thresholded before the LLM is allowed to answer. Distinct from the AizenvoltPrime Damocles listing. Project guide.
• Jev Decides, Claude Writes (Claude Code hook that delegates each message to Haiku, Sonnet, Opus or Fable) - Claude Code UserPromptSubmit hook: Jev answers one Choice question about each message (the smallest model that can do the job) and the main session delegates to one of four helper agents; short follow-ups, low-confidence picks, slow answers and messages that look like they contain a secret stay with the main session. Project guide.
• Jev Decision - Portable, budgeted Jev client for coding-agent harnesses (Claude Code, Codex, Cursor, Gemini CLI, OpenCode, any MCP or shell client) with a shared daily budget ledger, explicit unavailable results and harness recipes; advice, never authority. Project guide.
• Jev Decision Gateway (kartikanand73) - Governed withdrawal PoC: TypeSafe Jev scores, deterministic policy signs, executor only accepts signed decisions; bench vs GPT-6 Sol and rules — distinct from kuldeepsinh19/jev-decision-gateway. Project guide.
• JEV Decision Hub - Zero-dependency Python TypeSafe Jev client with key-pool rotation, a small web dashboard, and ready examples (moderation, routing, trades). Project guide.
• Jev decision-model benchmarks (vilinski) - Reproducible decision-model benchmark: a hand-labelled own set (60 tasks x3) and a 1,100-item public sample, run against Jev and other System One routes, with raw results, pricing snapshots and generators for a published report and interactive model explorer. Project guide.
• Jev Decisions Plugin for Hermes - Hermes (and other agent) plugin adding Jev-backed tool-risk reviews, approval recommendations, source checks, and task-completion checks. Project guide.
• Jev Deep Research - Research agent pattern: GPT plans searches; TypeSafe Jev judges document regions in parallel and returns original evidence passages. Project guide.
• Jev DevTools (React inspector for Jev requests) - Drop-in React inspector for TypeSafe Jev, inspired by TanStack Query Devtools: a drawer that records your app's Jev requests and shows animated flows from context to questions to answers, choice probability bars, score and noul values, request history, retries and raw payloads. Project guide.
• Jev Engineering Cookbook - Curriculum of Jupyter notebook recipes for building with Jev one bounded decision at a time — state in, typed choice / noul / score questions out — with offline fixture runs that need no API key, a live mode, and explicit provenance rules separating synthetic pipeline checks from measured Jev results. Project guide.
• Jev evidence triage (screens PubMed abstracts for a systematic review with Jev, rules and confidence gates) - Demo pipeline that screens life-science abstracts for one systematic-review question (can faecal microbiota transplantation or a live bacterial product improve immune checkpoint inhibitor outcomes in cancer?). Plain code rules run first, Jev answers six questions per abstract in one call, and code confidence gates route each record to include, exclude or human review; an optional LLM pre-read only helps the human queue. Includes 24 fictional sample abstracts with planted traps, gold labels, evaluation with threshold sweeps and an offline mock. Project guide.
• Jev Filter (Apixly) - Semantic filtering for AI agent tools: npm package @apixly/jev-filter (CLI, Python library, MCP server) turns large tool outputs, retrieved records, code and logs into compact typed TypeSafe Jev decisions with source IDs, local originals and receipts before they enter an agent's context; batched requests and maintainer-run A/B benchmarks. Project guide.
• JEV Flaky Detective - GitHub Action: TypeSafe Jev classifies failing tests as regression/flaky/environment/unknown (never auto-reruns). Project guide.
• Jev Flow - Standalone Node.js studio for typed Jev workflows (Studio, Compendium, Battle Arena; distinct from jev-flow-harness). Project guide.
• Jev for Claude - Dependency-free Claude Code plugin with nine user-invoked engineering skills (safety preflight, rule rubric, browser QA, evidence completion and more), conservative local hooks, and an explicit opt-in typed TypeSafe Jev judge adapter. Project guide.
• Jev for Claude Code (brookcs3) - Unofficial Claude Code plugin exposing TypeSafe Jev typed judgments (noul/choice/score) plus corpus→eval loops. Project guide.
• Jev for Claude Code (pachelbel23) - Unofficial one-file installer that adds Jev hooks to Claude Code: skill suggestion and task-type classification per prompt, optional subagent model tier choice and shell risk confirmation, plus a jev-vision CLI for image yes/no, choice and Box questions. Project guide.
• jev for DuckDB (judoaseeta) - DuckDB port of pg-jev: ask TypeSafe Jev yes/no, probability, choice and score questions about table rows from SQL (distinct from prasanthj/duckdb-jev). Project guide.
• Jev for Excel - Unofficial Excel add-in that calls TypeSafe Jev from worksheet formulas — =JEV.ASK("choice"|"score"|"noul", A2, …), shortcut functions and JEV.STATE for multi-field rows — as an async Office add-in for Microsoft 365 or a Windows VBA edition. Project guide.
• Jev for Home Assistant - Turns judgments over selected Home Assistant state into sensors and automation responses, with usage accounting. Project guide.
• Jev for Marketers (free guide, Claude skill and recipes for marketing-ops decisions with Jev) - Free, hands-on guide to using Jev for the small yes/no calls in marketing ops (route this lead, tag this asset, is this reply interested), plus a Claude skill with four editable recipes (lead routing, reply triage, on-message check, asset tagging), a no-install batch script that runs a recipe over a CSV, and a 50-lead sample with an answer key for scoring. Project guide.
• Jev for Splunk - Splunk app adding a jev search command for noul/choice/score over events via TypeSafe Jev with KV-store cache (Apache-2.0 LICENSE; GitHub SPDX unset). Project guide.
• Jev Foundry Judge - Drop-in Azure AI Foundry agent evaluators powered by TypeSafe Jev, plus a BYOK demo app and model-router bench. Project guide.
• Jev Game Engine - Early Rust desktop prototype for controlling games through bounded goals: TypeSafe Jev selects a goal, a local executor handles movement, validation and stopping, and every decision is traced, timed and replayable; Minecraft Java 26.2 is the first integration, with an offline fixture preview. Project guide.
• Jev Gate (lymagics) - GitHub Action/CLI that asks TypeSafe Jev yes/no quality questions over selected scopes and fails the pipeline on threshold breaches (distinct from other listed jev-gate projects). Project guide.
• Jev Gate (royalskynet) - MCP server and Claude Code/Codex plugin that prunes tool output, compacts conversations and gives task/review advice with TypeSafe Jev; optional local Laya pilot (distinct from other jev-gate projects). Project guide.
• Jev Gatehouse (Kinde) - Kinde identity + permissions with TypeSafe Jev typed gates before every MCP tool call. Project guide.
• Jev GitHub Action - GitHub Action: install pinned Jev CLI, run typed judgments on an event/JSON request, expose answers (no issue mutation). Project guide.
• Jev Graph Search - CLI and agent skill that retrieves evidence from local Obsidian vaults, Logseq folders, or JSON memory snapshots: a local shortlist is reranked by Jev (TypeSafe or OpenRouter). Project guide.
• Jev Harness (Apixly) - Python runtime for open-ended goals: a primary AI defines tools, parameter providers and completion checks, then TypeSafe Jev chooses the next tool and arguments from the goal, history and fresh state while code executes, verifies and enforces bounds. Install from Git; the PyPI name jev-harness is an unrelated project. Project guide.
• Jev Highlight Cutter - Codex agent skill and stdlib Python CLI (Chinese docs) that cuts long interviews, podcasts and streams into an evidence-backed highlight reel of a target length: transcript windows are scored by Jev against your preferences, selections are traceable, and FFmpeg makes a local rough cut. Project guide.
• Jev Highlight for DeepSeek Harness (highlight the parts of long AI answers worth reading) - Web plugin for DeepSeek Harness that leaves AI answers unchanged but highlights what to read first: it splits an answer into units, asks a decision model four yes/no questions about each (conclusion, action, risk, relevance) and shades them by priority; StartLux-Decision is the default model and TypeSafe Jev is a supported option. Chinese-language docs. Project guide.
• Jev IDS - Few-shot network intrusion detection with Jev, benchmarked on NSL-KDD against an LLM and classic ML. Project guide.
• Jev Improv Piano (a piano that composes and plays bar by bar, with Jev choosing each bar) - Static web piano that turns a one-line idea into a piece and plays it live: once per bar, Jev answers eight questions in parallel (harmonic function, chord, left-hand texture, rhythm, direction, development, intensity, whether to breathe) from a music-theory-legal candidate set, and a deterministic engine renders the notes. Plays offline with a seeded local decider when there is no key. Modelled on cocktailpeanut/jevthoven; Chinese-language docs. Project guide.
• Jev in Large Language Gibbs - Replication of the TruthfulQA consistent-reasoning experiment from Large Language Gibbs (arXiv 2606.19264) with Jev supplying per-claim conditionals, plus controls, GSM8K, knights-and-knaves, entity resolution and kinship tests of where Gibbs helps. Project guide.
• Jev Internal Links - Claude Code skill: crawl a site, ask TypeSafe Jev which internal links to add, apply local rules, export a report. Project guide.
• Jev Java SDK (dmishra2022) - Community Java 26 client for the System One API with typed noul, choice and score questions, retries, timeouts and virtual threads, plus an optional structured-concurrency module and Spring Boot 4 sample. Project guide.
• Jev kit by Mr-heka (record browser or Mac app chores and replay them with Jev) - Skill for Claude Code and Codex: the agent records a browser or Mac app chore into a recipe once, then a runner replays it at machine speed, sending only fuzzy calls (right page, right result, which option fits) to Jev through OpenRouter. Project guide.
• Jev Lab - Local Hundred NPC-town and Jev Shogi labs where TypeSafe Jev chooses the next legal action (Rules mode without a key). Project guide.
• Jev Lab by Nijal AI (45-minute French workshop app that pits Jev against an LLM on speed, cost and calibration) - Web workshop for a room of participants on phones or laptops, no account or code: try Jev and an LLM on support, scam and moderation scenarios, play a scam-filter mission where you set allow/review/block limits on 22 fictional messages, then a duel on errors, latency, cost per 1,000 decisions and calibration. Facilitator tab with a leaderboard and a cache-only fallback. French UI; works in simulated mode without keys. Project guide.
• Jev LLM Quality Tester - Prototype dashboard and FastAPI service that scores LLM-generated answers against explicit criteria with a Jev/TypeSafe structured-scoring adapter, SQLite history and a no-key deterministic demo mode. Project guide.
• Jev LLM router - Python lab that asks TypeSafe Jev to route a query to an OpenAI model, then has a local Ollama judge review the same choice. Project guide.
• Jev Logs - Scores OpenTelemetry logs for a separate analysis branch, with mock mode and conservative error handling. Project guide.
• Jev Lucid Memory - Research prototype of wake–sleep agent memory: an LLM drafts lessons, a decision model (TypeSafe Jev via the official SDK or OpenRouter, or local Clef-Flash) decides what to keep and when a lesson applies, and code enforces evidence and scope. Project guide.
• JEV Mail Filtering - Local read-only IMAP inbox triage with TypeSafe Jev (needs reply / worth reading / scam / unsure). Project guide.
• Jev Mario by imink (Super Mario Bros played by Jev decisions, with a full-campaign replay site) - Super Mario Bros in a real NES emulator where structured game state goes to a Jev Choice that picks the controller input; a browser replay site plays back all 32 stages and 6,275 recorded decisions with their answers, without calling a model; bring your own legally obtained ROM. Project guide.
• Jev MCP (asimons81) - Small community MCP server that exposes three bounded Jev decision tools (yes/no, choose and score) over stdio or authenticated loopback HTTP, with text-only inputs, no search or action execution, and credentials read only from the environment. Project guide.
• Jev MCP (Freepik) - Go MCP server for typed decide/classify/verify/rerank with Jev via OpenRouter or TypeSafe (binary/container releases). Project guide.
• Jev MCP (tmbmartell) - Portuguese-language MCP server that exposes Jev (TypeSafe System One) to agents such as Hermes, Claude Code, and Cursor as a typed judgment tool for noul, choice, and score questions, routed via OpenRouter or TypeSafe. Project guide.
• Jev MCP Server (keysersoft) - AnythingMCP connector exposing TypeSafe Jev yes/no, classification, and scoring tools to Claude, ChatGPT, and other MCP hosts (hosted or self-hosted). Project guide.
• JEV MLX - Typed decisions from existing MLX-LM and MLX-VLM models on Apple Silicon without fine-tuning or extra decision weights. Project guide.
• Jev Mode - Claude Code skills and study course for building with TypeSafe Jev (PolyForm Noncommercial; not all skills call Jev). Project guide.
• Jev Model Router - Early-access Claude Code mod that uses Jev task assessments and configurable confidence thresholds to route subagent models and main-conversation reasoning effort. Project guide.
• Jev Model Router (Codex skill) - Codex skill (Chinese and English docs) where the parent session condenses a non-trivial task and asks TypeSafe Jev to recommend the model tier and reasoning effort for an execution subagent; a standard-library Python script validates the answer against local policy and outputs one JSON decision, keeping the current model on timeout or invalid responses. Project guide.
• Jev Model Router by Akamai Developers (WebAssembly model router with Spin and TypeScript, Jev picking the model) - Sample from Akamai's developer team: a single Spin WebAssembly component exposing POST /ask that asks Jev (via OpenRouter) which model should handle each prompt, forwards the prompt to that model on an Ollama instance and returns the answer. Accompanies the blog post "Build a Model Router with Jev, Spin & TypeScript". Project guide.
• Jev Model Router by zerone-agents (OpenAI- and Anthropic-compatible gateway where Jev picks the model) - Go model-routing gateway for agents: OpenAI Chat Completions and Anthropic Messages clients connect to it, and either name a model or let Jev choose from editable model cards; it streams text and tool calls, stores configuration in SQLite and ships a CLI, bilingual dashboard, Docker Compose quickstart and an agent skill. Distinct from the separately listed Claude Code and Codex routers of the same name. Project guide.
• Jev Model Routing Lab - TypeScript lab: TypeSafe Jev typed signals drive confidence-aware Claude/Kimi model routing with simulated offline demos. Project guide.
• Jev Moderation - TypeScript and Python SDKs for text moderation with TypeSafe Jev: a bundled, versioned community policy or your own filters, each rule compiled to a Noul question, with review/block thresholds and a policy test CLI (validate, eval, compare with quality gates). Project guide.
• Jev Moderation API - Small FastAPI proof of concept that moderates text against configurable policies with Jev Noul questions via the official TypeSafe Python SDK. Project guide.
• Jev Moderation for Laravel (confidence-aware content moderation) - Laravel package that classifies any Eloquent model with Jev when it is saved and uses per-category probability thresholds to pass it, hold it for human review or block it, keeping an append-only history of every decision and review. Project guide.
• Jev NetHack - TypeSafe Jev plays NetHack 5.0: code lists legal moves, Jev picks one (dashboard + optional Hardfought). Project guide.
• Jev Observer - Local proxy and dashboard for TypeSafe Jev decisions (history, versions, latency/cost) with an embedded web UI. Project guide.
• JEV OpenFast Browser - Local chat-driven browser agent built on Jev Ultrafast: Jev (via OpenRouter) picks each operation and target from a numbered element table in a real Chrome tab, checks results before reporting success, and hands over for logins or CAPTCHAs. Project guide.
• Jev Orchestrator - Small Node/TypeScript experiment testing whether Jev can make useful next-step decisions around a coding agent: decision-only by default (inspect the repo, send compact state to Jev via TypeSafe, Vercel or OpenRouter, apply deterministic thresholds, record a JSONL trace), with an explicit --orchestrate mode for a bounded, manually approved loop of safe searches, reads, read-only git, validation scripts and delegation to Codex CLI or Claude Code. Project guide.
• JEV Orchestrator (Claude Code plugin where Jev picks the model, effort and tools for every subagent) - Claude Code plugin that intercepts every Agent spawn and asks Jev which model (Haiku, Sonnet or Opus), effort level and tool profile the subagent needs, plus a /jev-orchestrator:run command that executes a prompt file wave by wave with parallel subagents routed the same way. Failed tasks stop only the waves that depend on them, and commits happen only when you ask. Supports a TypeSafe key or an OpenRouter key, stored in the system keychain. Project guide.
• Jev Outlook Cleaner - Local web app that triages older Outlook Inbox messages into keep, review and delete candidates with Jev plus local mail signals, auto-moves only high-confidence bulk mail to Deleted Items, and never deletes permanently. Project guide.
• Jev Paired-Replay Study - Pre-registered study of whether identical replays flip Jev 1.13's action at shipped thresholds: 12,705 calls on 847 near-threshold items from Banking77, CLINC150 and HelpSteer2 found about 18.5% flips at 0.5 and 0.6 and 5.3% at 0.9, with flips nearly cancelling out. Project guide.
• Jev Playground (xingqiux) - Local, standard-library Python test bench for Jev: enter text, JSON or images plus Noul, Choice, Score or Box questions, run one request, and see probabilities, raw request/response, cURL and a choice-order sensitivity check, with 60 built-in examples. Project guide.
• Jev Playground by gmassello (step-by-step Spanish-language Jev lessons via Vercel AI Gateway) - Small local web playground, written in Spanish, that teaches Jev step by step through the Vercel AI Gateway: you try typed questions, see confidence and per-call cost, and read practical notes on calibrating thresholds, measuring limits and keeping the key server-side; it also compares the native TypeSafe API and Cloudflare Workers AI routes. Project guide.
• jev plays snake - Browser Snake where code computes legal moves and exact facts (food distance, flood-fill reachable cells, dead ends) and TypeSafe Jev picks one move per tick with a single choice question through a small Hono proxy; misses the deadline → the snake goes straight. Project guide.
• Jev RAG Chinese Benchmark - Reproducible Chinese RAG reranking benchmark for TypeSafe Jev (T2Reranking, pinned model, cost caps, per-call logs). Project guide.
• JEV Reasoning Navigator - Cognitive supervision runtime: TypeSafe Jev semantic judgment plus policy engine, capability receipts, and sandboxed tool execution (license unspecified). Project guide.
• JEV Research MCP - Local-first MCP research tool: TypeSafe Jev filters search hits and page blocks before dumping them into an agent context. Project guide.
• Jev resume screener (Claude Code plugin) - Claude Code plugin that screens a folder of resumes against a job description with Jev: fit ranking, copied-resume clusters, flags for resumes mirroring the job description, and a human review queue. The output is a sorted reading list, never a rejection. Project guide.
• Jev Review - MCP server for experimental software-quality rubric scores and comparisons; requires JEV_API_KEY and sends supplied code context to TypeSafe. Project guide.
• Jev Review (Dev Agrawal) - Screens JavaScript and TypeScript changes or source files with staged Jev judgments and displays evidence-linked findings in a local dashboard. Project guide.
• Jev Review Action - Configurable GitHub Action: catalog or PR classification with TypeSafe Jev only (no text-gen), one template comment. Project guide.
• Jev review gate - GitHub Actions gate: local rules + TypeSafe Jev questions on the diff; exits 0/1 (does not merge). Project guide.
• Jev Review Markers for Premiere - Premiere UXP panel that turns typed Jev review findings on an SRT caption export into timeline markers: Jev flags issues and picks the exact caption, and every marker is added in one undoable transaction with the caption text copied byte for byte. Project guide.
• JEV robot (Unitree G1 in MuJoCo with Jev motor decisions and Gemini planning) - Simulation-only demo of a Unitree G1 humanoid in MuJoCo: Jev chooses velocities and bounded joint offsets (full control) or named movements (primitives) while Gemini plans from the live camera frame, all shown in a browser control room; an offline mode runs without keys. Project guide.
• Jev Router by AIAnytime (OpenAI-compatible router where Jev picks the model, agent or tool for each request) - Self-hosted, OpenAI-compatible endpoint: send model="auto" and Jev's typed choice, score and yes/no answers decide which configured model, agent or tool handles the request, with the routing reason returned alongside the response and a dashboard for traces. Project guide.
• Jev Router for Windows - One-click Windows GUI for TypeSafe Jev routing with Codex (jev-codex-bridge) and assisted Claude Code plugin setup. Project guide.
• Jev RPA Assistant (Windows customer-service prototype that picks approved replies with Jev; Chinese) - Windows desktop prototype (Chinese README) for e-commerce customer service: Clicknium RPA reads the chat page, a vision model transcribes it, and Jev chooses step by step through an approved reply library or hands the chat to a person, with page-state checks before and after sending and a SQLite audit trail. Adapted for Xianyu web in Edge; 1688 is a later target. Project guide.
• Jev Rules for WordPress - Experimental WordPress plugin that queues posts for Jev classification against versioned rule packs and keeps an administrator review history; posts are never changed, and other plugins can enqueue posts and react to decisions through hooks. Project guide.
• Jev Runway - Local Codex proxy that uses TypeSafe Jev to trim stale tool output and save tokens. Project guide.
• Jev Score - Local CLI/web scoreboard that grades document revisions against your criteria with TypeSafe Jev via OpenRouter Decisions. Project guide.
• Jev Second Brain - Local-first Markdown vault CLI: FTS index, source-linked related-note suggestions, optional TypeSafe Jev judgments via Vercel AI Gateway. Project guide.
• JEV Sees - Give TypeSafe Jev eyes: one call, many object judgments from camera/video frames. Project guide.
• JEV Shield by JamesANZ (Jev-scored prompt-injection and command gate for coding agents) - Guard that sits outside the model: it scores incoming prompts and pasted text for injection with Jev, and checks proposed shell commands (local rules first, Jev for the rest), then ordinary code maps the probabilities to allow, ask or block, including as a Claude Code permission hook; the README walks through a Gandalf password-game test. Project guide.
• Jev Sift - Screens files, public webpages, and tool descriptions with Jev before an agent reads selected content; requires a TypeSafe key, and upstream licensing is unspecified. Project guide.
• Jev Skill Router (kelasdev) - OpenCode plugin that uses TypeSafe Jev (two gated /v1/systemone calls) to inject at most one relevant SKILL.md into the system prompt. Project guide.
• Jev Skills (n23eos) - Fourteen installable Claude Code and Codex skills plus a CLI where TypeSafe Jev makes one bounded choice (skill, model, context file, first test, bug component or plan) and the coding agent does the work; distinct from the listed jev-skills plugin. Project guide.
• Jev Skills Retrieval - Pi coding agent extension that retrieves skills per request: Jev scores skill metadata once per turn, and Pi then reads only the selected skills' bodies, with pinned skills and explicit /skill:name taking priority. Project guide.
• JEV SOC alert triage (classifies exported Windows security events with Jev and routes uncertain ones to review) - Python CLI that classifies exported Windows security events, assigns investigation priority and routes uncertain or high-risk results into a review queue, writing one JSONL decision trail per event (detector hits, backend and model, confidence, severity, routing reasons). Jev is the model backend; a rule baseline runs without credentials and an optional Ollama stage writes investigation notes. Project guide.
• Jev Starter - Express + React starter for small web apps that make Jev decisions via OpenRouter. Project guide.
• Jev Starter - Browser console to configure TypeSafe Jev Noul/Choice/Score questions, test live, and export Python/JS/cURL (or example JSON). Project guide.
• Jev Support Ticket Router - Support-ticket routing demo (English/Chinese) that asks Jev department, urgency and tone in one call and sends low-confidence tickets to a human queue. Project guide.
• Jev Symfony Bundle - Unofficial Symfony bundle for TypeSafe Jev: typed client, validator constraints, Messenger, Workflow guards, and profiler. Project guide.
• Jev Talking Head Workflow (talking-head video production skills) - Reusable workflow of agent skills for Chinese talking-head videos, from script to rough cut, B-roll, subtitles, mixing and QA, where Jev is the structured review layer comparing takes, flagging fragments and repetition and suggesting layouts, while humans confirm every edit. Project guide.
• Jev Task Router - npm CLI that asks Jev, through the AI SDK and Vercel AI Gateway, which eligible coding model should handle a task, then hands the task to your local Codex or GitHub Copilot CLI session after you confirm. Project guide.
• Jev the Janitor - Markdown vault janitor: TypeSafe Jev votes on each note; code adds frontmatter or quarantines secrets (dry-run default; offline mode). Project guide.
• JEV ToC search (search a PDF through its own table of contents, with Jev ordering each branch of a depth-first walk) - Experiment in long-document search inspired by PageIndex: parse a PDF with Datalab, recover its real table of contents and match printed page numbers to PDF pages, then let Jev order the branches at each level of a depth-first search. Low-probability branches are never pruned, and in the worst case every page is checked before it says no evidence was found. Includes grounded answer evaluation and interactive traces. Project guide.
• Jev Trader - Studies Jev market-direction choices, simulated fills, and on-chain order execution through a Bun trading experiment. Project guide.
• Jev Trader by Errr0rr404 (local trading desk where Jev votes and code owns sizing and exits, simulate by default) - Local CLI and web desk that uses Robinhood's Agentic MCP for quotes and orders, asks Jev to vote on entries, and leaves sizing, stops and scale-outs to code. Simulates a $100 paper book by default; live orders need an explicit armed mode. Project guide.
• JEV Trading Command Center - Paper-trading research app (README in Traditional Chinese) where Jev decisions pass validation and risk checks before a simulated broker acts, shown in an animated 3D command center with model 'pets' and an offline demo that uses only synthetic data. Project guide.
• Jev Ultrafast - Selects browser operations and observed targets with Jev, using a separate text model for text entry. Project guide.
• Jev Unpack - Codex/Claude Code Stop hook where Jev detects multi-topic, multi-ask agent messages and makes the agent pace them one ask at a time. Project guide.
• Jev Valley (simulated valley where rabbits, deer and wolves decide with Jev) - Small ecosystem simulation where every animal picks its next action once per valley hour with one Choice question and a few Score and Noul questions, watched live in a 3D browser viewer, with replays and month-long result notes. Project guide.
• Jev Vampire Survivors - TypeSafe Jev picks character, stage, level-ups, and walk direction in real Steam Vampire Survivors. Project guide.
• Jev vehicle classifier (Korean 12-class traffic survey) - Episode 1 of a Korean ITS series: vision-style features (axles, units, trailer type, load tons) become a JSON state and Jev picks one of the 12 official Korean road-traffic vehicle classes with probabilities; low confidence goes to an unclassified bucket for review (README in Korean). Project guide.
• Jev Video Lab (ask structured questions about video footage with Jev and evaluate it on annotated clips) - Experimental local app for asking video questions with Jev: event search across indexed videos (before, after, overlap), multiple-choice questions on a segment, object presence and Box locations drawn on frames, procedure checks that flag missed steps for a person, recording alignment, human review and exports. The author's first study on Perception Test reports 463 of 658 questions right (70.4%) and weak precise event localization. Project guide.
• Jev Vision Benchmarks - Benchmarks of Jev 1.14 image input on marketplace tasks (product-photo moderation, photo-to-listing matching, classification, package-label reading) against GPT vision models, AWS Rekognition and OpenAI moderation, with checked-in labels and results. Project guide.
• Jev Voice Browser - Experiments with typed Jev decisions over partial voice transcripts and page elements to control Playwright, with a local decision inspector. Project guide.
• Jev vs Basal insurance benchmark - Polish motor-insurance (autocasco) decision benchmark: 325 synthetic typed decisions built from public policy terms, with code-computed labels, run against Jev, Basal 2.0 and Gemini models with accuracy, calibration (ECE) and latency. Project guide.
• Jev vs Parallel Decisions (two real Jev gates scored against answer keys fixed before either tool ran) - Benchmark of Jev against Parallel's new web-searching Decisions API on the questions two of Eric Siu's production Jev gates ask: M&A and ICP criteria for 50 YC companies, and whether a content topic has already been published. Answer keys come from the YC directory and the site's own pages, fixed before either tool ran; results are reported as balanced accuracy. Project guide.
• Jev vs. LLM (Banking77) - Small, fair test of the "up to 200x faster" claim: 100 real Banking77 customer messages sorted into 77 categories by Jev, Claude Haiku 5.5 and Claude Fable 5.1, measuring accuracy, off-list answers, median and p90 latency, and cost. Project guide.
• Jev WCAG Auditor - axe-core WCAG checks plus TypeSafe Jev adjudication of judgement calls on public/.gov pages. Project guide.
• JEV WHERE? - Claude Code plugin that finds where Jev could make decisions in your app and writes the typed Choice/Score/Noul question for each (analysis only). Project guide.
• Jev Workflow by cathy-kim (Claude Code plugin and guide for designing Jev decisions offline) - Community toolkit for designing typed decision workflows: a portable skill with Claude Code, Codex and Hermes adapters, offline examples, an illustrated guide, and an opt-in live client and stdio MCP server for TypeSafe System One, OpenAI Decisions and OpenRouter. Project guide.
• Jev Workflows for Codex and Claude Code - Open-source Codex and Claude Code plugin that consults Jev about tools, models, tasks, skills, context and custom decisions, diagnoses failed commands and checks completion claims against evidence; eight MCP tools, lifecycle hooks and three standalone skills. Project guide.
• Jev Workshop (notebooks and deck) - Teaching kit for Jev: five notebooks executed live against the API, a 42-slide deck with speaker notes, and a cited reference covering the primitives, confidence and common patterns. Project guide.
• Jev Workspace (give Jev a mission and some links; it judges every item live and hands you a results page) - Local app: type a goal, paste links or drop files (sites, sitemaps, feeds, CSV or Google Sheets, PDFs, Instagram exports), let Claude draft a yes/no gate plus breakdown questions, and watch Jev judge every item live with counters, throughput and cost. Matches, rejects and an uncertain band that Claude reviews; filters, CSV/JSON export and an optional Mac app wrapper. Project guide.
• Jev x AI SDK Form Router - Next.js form router that uses Jev to route submissions by context, with an LLM fallback for uncertain cases. Project guide.
• Jev × LexGLUE (reproducible zero-shot benchmark of Jev on all seven LexGLUE legal tasks) - Benchmark that freezes the official LexGLUE evaluation sets, sends each example to Jev as typed choice or per-label yes/no questions through OpenRouter, checkpoints every request and reports micro/macro-F1, usage and cost. The README compares Jev 1.13 with a frontier chat model, open System One models and a fine-tuned BERT baseline on the full 23,607-example test split. Project guide.
• jev-a2a (Jev router for Paseo agents) - Personal-prototype router that sends work between coding agents running in Paseo sessions on several machines and keeps one record of what was asked, who took it and what came back; when a request names no recipient, TypeSafe Jev chooses the participant. Includes a web board and design docs; published as a worked example, not a product. Project guide.
• jev-adb - MCP server for Android control over ADB where TypeSafe Jev makes typed UI decisions behind consent and safety gates. Project guide.
• jev-advisor (gjc-plugin) - gajae-code (gjc) plugin set whose jev-advisor asks TypeSafe Jev for fast/balanced/strong tier advice on subagent delegations, in shadow mode by default, with offline reports. Project guide.
• jev-advisor-gateway - Local gateway that routes coding-agent tool decisions through Jev and can attach a read-only advisor (GPT or Claude) without giving the advisor write authority. Project guide.
• jev-agent-browser - Confidence-gated browser actions for agent-browser with TypeSafe Jev (or Gateway/Cloudflare/custom providers). Project guide.
• jev-agent-failure-benchmark - Benchmarks TypeSafe Jev on Who&When Pro agent-failure attribution (who/when/what) against published LLM baselines. Project guide.
• jev-agent-kit - Zero-dependency CLI + MCP tools (route/triage/guard/grep/rank/compact/judge/…) on TypeSafe Jev (@walidboulanouar/jevkit; distinct from Rust jevkit). Project guide.
• jev-agent-kit (nanoDBA) - Evidence-layer hooks for Claude Code/Codex/Hermes: TypeSafe Jev scores risky tool calls in shadow mode by default — distinct from walidboulanouar/jev-agent-kit. Project guide.
• jev-agent-kit by bitskc (self-hosted Jev MCP toolbox and skills for oh-my-pi and Claude Code) - One Cloudflare Worker exposing 15 Jev-backed MCP tools, plus 7 agent skills and an optional guard hook, for the judgment calls coding agents make all day: is this shell command safe, is this sub-agent the right size, did the agent really verify before saying done. Builds on jkudish/jev-mcp and jonathanavis96/jev-kit; deterministic rules handle what a model shouldn't guess. Project guide.
• jev-agent-memory (no vector DB: Jev picks which past chats and profile facts a Claude Code agent recalls) - Unattended memory for a Claude Code agent you talk to through Telegram or Discord: hooks capture the chat, Claude Haiku summarizes it into daily Markdown logs every 30 minutes, and on each message Jev picks the topic and which past excerpts would help, injecting at most four. A daily job distills up to three evidence-backed traits into a user profile. Unofficial; English and Korean docs. Project guide.
• jev-agent-router - Claude Code mod that picks the model and effort of every subagent from its task type, with Jev as an optional classifier. Project guide.
• jev-agent-skill (RosarioDiBartolo) - Agent skill connecting Jev and Kev System One models to Codex and AI agents for fast tool routing, classification, and bounded decisions. Project guide.
• jev-agent-tools - Jev transport/provider layer with fail-closed validation; shared by jkudish/jev-browser and jkudish/jev-mcp. Project guide.
• jev-ai-demo (Jev primitives walkthrough) - Small Python walkthrough of Jev's Choice, Score and Noul primitives: a quickstart, atomic scoring and confidence routing examples that replay the sample responses from TypeSafe's docs without a key, then call the live API when one is set. Project guide.
• Jev-AI-Skill - Claude Code/Codex/Hermes skill + MCP: TypeSafe Jev gates for agent decisions. Project guide.
• jev-ai-use-cases (atliq) - LangChain notebook: TypeSafe Jev for triage, routing, reply guards, tool selection, finance-inbox checks. Project guide.
• jev-align - CLI to build calibrated AI Functions from human labels using TypeSafe Jev evaluation and GEPA optimization. Project guide.
• jev-android - Kotlin Android UI-agent SDK: TypeSafe Jev or DeepSeek selects accessibility actions; local Maven sample (not on Maven Central). Project guide.
• jev-answers - MCP server + CLI: send files and typed questions to TypeSafe Jev and get the raw answer back (one session folder per request). Project guide.
• jev-apply - Job-application form filler: rules plus a decision backend (local Laya default, TypeSafe Jev selectable) choose facts/options from your profile; stops before Submit. Project guide.
• jev-audit (agent skill) - Agent skill that audits a codebase for brittle fact-guessing rules and writes a ranked JEV_OPPORTUNITIES.md with evidence, fit scores, a replacement sketch that keeps reliable checks in code, and a measurement plan; makes no Jev calls itself. Project guide.
• jev-auto-approve - Claude Code hook: TypeSafe Jev auto-approves read-only shell commands in milliseconds; everything else still prompts you (fail-open). Project guide.
• jev-auto-mode (OpenCode plugin where Jev judges permission requests) - OpenCode V2 plugin that adds an auto mode: edits are approved, and Jev judges other permission requests for task relevance, data destruction, secret exposure, data transmission and unauthorized remote or system changes, denying anything uncertain. Project guide.
• jev-autopilot - Claude Code plugin that keeps unattended sessions moving: TypeSafe Jev answers Claude's routine questions, screens every non-read-only tool call, and holds risky ones until you tap Proceed or Block on your phone via Telegram; falls back to the terminal when Jev is unreachable. Project guide.
• jev-backed-function-calling - Thin Python layer that uses TypeSafe Jev as a verifier behind function-calling flows (TYPESAFE_API_KEY / JEV_API_KEY), so tool arguments can be checked with typed System One questions. Project guide.
• jev-backend-qa - Backend QA agent with PAL/Jev adjudication to BLOCK/WARN/PASS (CLI + GitHub Action). Project guide.
• jev-bench (stern9) - Benchmark harness that asks Jev and general LLMs (via system-one-adapter) the same routing, yes/no and rating questions and reports accuracy, p50/p95 latency, calibration and cost. Project guide.
• jev-blindspot - Claude Code / Codex side panel: TypeSafe Jev gates whether a prompt needs a second look, then optional blind-spot cards (issue #400). Project guide.
• jev-box-tracking - Python library and webcam demo for real-time object tracking: Jev Box answers give the absolute position, optical flow tracks motion between answers, and a Kalman filter rewinds and fuses late answers. Project guide.
• jev-brain (one GPU serving chat and Jev-compatible decisions from one open model) - Self-hosted gateway in front of a single vLLM process that serves OpenAI-style chat and Jev-compatible /v1/systemone decisions at the same time from one loaded model, with nine profiles across Gemma 4 and Qwen models in several quantizations (some with image decisions) and published Decision Index and latency results; independent of hosted Jev. Project guide.
• jev-broker - Local HTTP MCP broker for Hermes agents: validate questions, call TypeSafe Jev via OpenRouter, return structured judgments (agent still decides). Project guide.
• jev-browse (cooper667) - Claude Code plugin: plain-English Playwright checklist QA judged by TypeSafe Jev on Cloudflare Workers AI (distinct from other jev-browse listings). Project guide.
• jev-browse (danielnc) - browser-harness helpers: TypeSafe Jev decides each click for agent-delegated website sub-tasks (distinct from cooper667). Project guide.
• jev-browsecomp - BrowseComp-Plus research harness: Jev Choice/Noul screening vs full-LLM/RLM pipelines with mechanical citation checks. Project guide.
• jev-browser by MaciejWiatr (Claude Code browser control in background Chrome tabs, with Jev Ultrafast runs) - Claude Code plugin that gives Claude the Claude in Chrome tool set over the Chrome DevTools Protocol in background tabs of your real, signed-in Chrome, with a separate agent cursor, a human-like mouse mode, per-task tab groups and remembered site actions, plus an optional jev_run tool that hands a narrow goal to the Jev Ultrafast agent. Project guide.
• jev-browser-bridge - Plug any CDP browser into Jev—cloud, local, or self-hosted, including no-render engines. Project guide.
• jev-browser-mcp (bothuany) - MCP browser: host states intent; TypeSafe Jev clicks; cheap reader sees the page; DOM stays out of context. Project guide.
• jev-browser-sidekick-mcp - MCP server: plain-language browser steps; TypeSafe/OpenRouter Jev picks page controls while Playwright MCP shares the Chrome session. Project guide.
• jev-browser-skill (ChenYCL) - Browser/computer-use CLI+MCP for coding agents: TypeSafe Jev chooses actions in ego lite/Chrome/Safari (distinct from zurfyx/hqman). Project guide.
• jev-browser-skill (hqman) - Playwright Chromium skill/jb CLI driven by Jev (Vercel AI Gateway or TypeSafe); distinct from zurfyx/jev-browser-skill. Project guide.
• jev-browser-use (AuroraPixel) - Jev browser loops for Codex/Claude: warm Chrome daemon, CLI/MCP/extension, host handoffs for reading/text. Project guide.
• jev-calibrate - Calibrate TypeSafe Jev questions against labelled examples and get per-question gate/ranker/unusable verdicts. Project guide.
• jev-calibration - Reliability diagram and ECE for TypeSafe Jev on 240 hand-labelled agent tool calls (labels + scripts published). Project guide.
• jev-calibration-audit (independent test of whether Jev's probabilities are calibrated, on US petroleum reports) - Independent audit of Jev's calibration on questions about EIA weekly petroleum inventory reports, where the right answer can be computed exactly: reliability diagrams, ECE, Murphy decomposition and a Spiegelhalter test, compared with an LLM classifier asked the same questions word for word, plus rules and climatology baselines. Project guide.
• jev-cascade - Python cascade that uses Jev as a cheap, recall-tuned first stage over many questions per item, then only escalates the uncertain band to a stronger second-stage command and reports whether stage two actually improves accuracy. Project guide.
• jev-cc-codex-router - Per-turn Codex model-routing proxy: TypeSafe Jev picks tier, rewrites the model, retries flaky upstream errors (distinct from 0xNatoshi/jev-codex-router). Project guide.
• jev-certify - Conformal risk certificates and prediction-powered audits for TypeSafe Jev routing thresholds (offline tests; live via OpenRouter Decisions). Project guide.
• jev-chess (byrencheema) - Experiment where Jev plays chess by choosing one move from all legal moves as a single Choice per turn, with no search or filtering: 350 games vs Stockfish and Maia, 50 vs random, 1,000 Lichess puzzles, Stockfish analysis and all results committed. Project guide.
• jev-chrome-mcp - MCP adapter running the jev-browser-use click loop in Chrome for Cursor/Codex. Project guide.
• jev-ci-pathfinder - GitHub Action: TypeSafe Jev selects which allowlisted CI jobs to run after a change; workflow YAML stays in control (distinct from jev-ci-selector). Project guide.
• jev-ci-selector - GitHub Action that asks TypeSafe Jev which described CI tasks apply to the current PR diff and exports boolean job outputs. Project guide.
• jev-ci-triage - Classify failing CI tests (Playwright/JUnit) as regression, flaky, environment, test-data, or unknown—rules first, TypeSafe Jev for ambiguous cases; never reruns or masks failures. Project guide.
• jev-classify-prompt (policy-based text moderation with Jev, as a Go library and CLI) - Go library and command-line tool, standard library only, that checks a text against a moderation policy: each class is one yes/no question, all go to Jev in one call, and a class matches when its probability reaches its line; a missing or malformed answer is an error, never a pass. Ships a ten-class default policy and started as the moderation in Human42. Project guide.
• jev-claude - Python library, CLI and MCP server that sends code-facing decisions (decide/when/switch/rank) to TypeSafe Jev and human-facing text to an LLM you choose, with low-confidence answers routed to a person. Project guide.
• jev-claude-code (Axiumine) - Research catalog plus reference MCP: TypeSafe Jev yes/no triage for Claude Code via OpenRouter (distinct from DarioFontanel/weiping). Project guide.
• jev-claude-code (DarioFontanel) - Claude Code prompt pack: TypeSafe Jev model/effort routing, tool-call compaction, and thresholded diff review. Project guide.
• jev-claude-code (weiping) - Claude Code plugin: TypeSafe Jev in the agent loop for permission gate, output ladder, conditional context, and subagent routing (shadow mode default; distinct from DarioFontanel/jev-claude-code). Project guide.
• jev-claude-gate (Claude Code hooks that ask Jev to vet session rules, tool calls, tool output and task completion) - Four Claude Code hooks backed by Jev: a SessionStart check of CLAUDE.md, AGENTS.md and rules files for planted instructions, a PreToolUse gate that scores each command or edit on several risk questions before it runs, a PostToolUse scan of fetched pages and files for indirect prompt injection, and a Stop check for agents that claim to be done while hiding errors. The gate sees the tool name, arguments and recent workspace state, not the agent's reasoning. README in Chinese. Project guide.
• jev-claude-plugin (Jev around Claude Code tool calls) - Claude Code plugin that puts Jev at four points around tool calls: it windows long tool output and withholds irrelevant or injected spans, can replace auto mode with a hazard-based allow/deny/ask check, suggests effort and model, and can check final claims against tool results. Project guide.
• jev-claude-router (Flam1ngFir3ball) - Claude Code plugin fork: TypeSafe Jev picks model tier and effort per turn with cost-aware switch checks. Project guide.
• jev-claude-router (sergeiboikov) - MCP server for Claude Code and Claude Desktop where TypeSafe Jev recommends which Claude model and effort level suit a task, comparing candidates by live OpenRouter prices, context sizes and output limits; it recommends only and never switches the model itself. Distinct from the separately listed jev-claude-router. Project guide.
• jev-claw - OpenClaw typed model routing with TypeSafe Jev (jev_route) and optional automatic enforcement. Project guide.
• jev-cli (portlandhodl) - Agent-facing Rust CLI over the unofficial jev-sdk crate: predict, choose and score subcommands return JSON probabilities an agent can gate on. Project guide.
• jev-cli (shaharia-lab) - Rust CLI/MCP for TypeSafe Jev: noul/choice/score with exit codes for shells/CI and JSON for scripts (distinct from tumf/jev-cli). Project guide.
• jev-cli (shetautnetjer) - Clean-room CLI for TypeSafe System One Decision Contracts, search, and benchmarks (distinct from tumf/shaharia-lab jev-cli). Project guide.
• jev-cli (tumf) - Unofficial PyPI CLI + stdio MCP for TypeSafe Jev noul/choice/score (distinct from okooo5km/jev, typesafe-cli, typesafeai-cli). Project guide.
• jev-click-bench - Benchmark of Jev 1.14's box tool as a GUI click model on ScreenSpot-v2 and ScreenSpot-Pro with the official scoring: 84% on ordinary screens, 19% on 4K pro apps, and 43% to 60% on Pro after one or two zoomed second looks, with every request logged. Project guide.
• jev-client (CLI, TUI, MCP server and web console for Jev through OpenCode) - Command-line client for asking Jev questions about a text or JSON payload through OpenCode's System One endpoint, with an interactive TUI, an MCP server so agents can call it, and a local web console; answers come back with probabilities per option. Project guide.
• jev-cloud-cost-guardian - GitHub Action FinOps gate: TypeSafe Jev evaluates cloud spend vs budget (approve/warn/block/manual-review) without mutating infra. Project guide.
• jev-cmdline-classifier - Agent skill/CLI that classifies shell commands with TypeSafe Jev Choice into allow/prompt/forbidden (advisory; sandbox stays authoritative). Project guide.
• jev-code-mode - Typed Jev judgments behind two MCP tools: search and execute. Project guide.
• jev-codex-router - Routes each Codex turn through Jev tier and thinking-depth selection via a local Codex Router generic provider, with fail-open fallbacks and an optional Codex-dry tandem. Project guide.
• Jev-Codex-Router by Emirhan Yüksel (Codex CLI hook that trims context per prompt with Jev) - Codex CLI UserPromptSubmit hook that asks Jev which rule areas a new prompt actually needs, leaves the others out of context and resets context for greetings and simple replies, with a hard 200 ms timeout and a local heuristic engine when there is no key or network; Turkish documentation. Project guide.
• jev-codex-token-saver - Codex plugin/MCP that scores bounded local evidence with TypeSafe Jev and returns selected exact excerpts (local fallback when Jev is unavailable). Project guide.
• jev-commit - pre-commit hook that uses one Jev call to check the commit message against the staged diff and flag leftovers. Project guide.
• jev-community-ops - Community triage for a developer community on TypeSafe Jev: typed questions, confidence-gated routing, weekly digest (simulate-friendly until early-access key). Project guide.
• jev-compact - Codex CLI plugin: TypeSafe Jev scores tool calls before compaction and restores critical outputs verbatim afterward (distinct from fast-jev-compaction for Claude Code). Project guide.
• jev-compaction (Waxmell114514) - Context compactor that can only score, never write—so agent memory cannot invent transcript facts (offline demo; distinct from fast-jev-compaction). Project guide.
• jev-compaction-plus - Claude Code compaction via TypeSafe Jev keep/drop (~0.5s): retained text stays verbatim; dropped tool outputs move to a drawer file (fork of fast-jev-compaction). Project guide.
• jev-compare (Jev image comparison: similarity score, differing regions and fix suggestions) - npm package that compares a candidate image with a reference using Jev and returns a 1–10 similarity score, the differing regions as boxes in both images, and each difference's type, element and fix suggestion in English or Chinese, with optional SVG outlines, saved JSON and HTML reports and request caching; built for design-vs-implementation UI review and visual regression checks. Project guide.
• jev-computer-use-skill (Mrchen116) - Codex/agent skill that delegates repetitive computer-use judgments to TypeSafe Jev. Project guide.
• jev-context-filter-harness - Paper code and run records for screening LLM agents' tool output with Jev: a relevance filter keeps chunks that answer the agent's question and an injection filter withholds planted instructions. Project guide.
• jev-controller - Fail-open advisory post-tool controller for OMP (oh-my-pi): asks TypeSafe Jev whether to inspect, edit, test, or ask the user—leaves the tool result unchanged if Jev is unavailable. Project guide.
• jev-cops - Context-aware tool-call policing for coding agents with graduated Jev-backed verdicts (allow→kill). Project guide.
• jev-corrective-rag - Corrective RAG demo: TypeSafe Jev typed gates for triage/grading/verify; generative LLM only writes answers. Project guide.
• jev-crowdsim - Test a message across an explicit grid of audience traits: Jev returns finite reactions (engage, trust, convert, main objection) for each declared combination, and code computes main effects and Wilson intervals instead of one average. Project guide.
• Jev-cu - Experimental Codex skill and JavaScript runtime for selecting macOS Accessibility targets with Jev, with action previews and optional result verification; execution-policy limitations require review. Project guide.
• jev-cu-win - Windows port of Jev-cu (docs in Japanese and English): a Codex skill where Jev reads the UI Automation text of a window and picks the next element to click or type into, with dry runs by default. Project guide.
• jev-datafusion - Rust crate that runs Jev decisions (noul, choice, score, ask) as Apache DataFusion SQL functions, so you can score tickets, route email or question meeting notes at warehouse scale. Project guide.
• jev-debtgate - Technical-debt gate for coding agents and CI: local metrics plus TypeSafe Jev typed questions and confidence policy (BYOK). Project guide.
• jev-decide (abdullahaamuda) - Agent skill + stdlib Python CLI: offload page-state Choice/Noul/Score judgments to TypeSafe Jev for browser/desktop automation. Project guide.
• jev-decision (gated-planner pattern where Jev may only pick one symbolic transition from a frozen whitelist) - Agent skill plus a dependency-free Python reference core for putting an LLM or Jev into a closed decision loop without letting it emit numbers, identifiers or verdicts. The planner picks one symbolic transition from a whitelist; deterministic code checks it against the failed gates, applies it to a copy of the state, preflights it and logs accept or reject. A snapshot script freezes the resolved Jev model ID before benchmarks. Project guide.
• jev-decision-harness - Claude Code plugin adding typed Jev decisions for command-risk gating, triage and review. Project guide.
• jev-decision-kit (kdcadmin) - Local skill/plugin cabinet: an on-device Jev head selects which skills to preface before the chat model speaks (distinct from DecisionKit .NET). Project guide.
• jev-decisions (wonghanz) - Typed decision layer for backend engineering (log triage, incident routing, PR triage, deploy risk) with a control-group eval harness via OpenRouter Jev. Project guide.
• jev-dev-demo (LIDR) - Spanish-language teaching repo with three runnable demos of 'Jev decides, the LLM reasons': Jira triage that classifies, prioritises and assigns each ticket with a confidence and escalates ambiguous ones, a pre-flight scan that flags existing bugs in the functions a ticket touches, and an LLM router that sends each PR to a large, small or no model (upstream reports 50–75% lower cost); every demo has a --mock mode. Project guide.
• jev-doc-search - Find answering pages in long PDFs with TypeSafe Jev Choice over a PageIndex tree (no vector DB). Project guide.
• jev-docs - Claude Code hook that asks Jev which pages in your project's docs folder matter for each prompt and adds a short reading list to the context; routine prompts get nothing. Includes a skill that writes docs in the format the hook reads best. Project guide.
• jev-dom - TypeSafe Jev operates any web page through its DOM action space (no WebMCP required); Playwright peer. Project guide.
• jev-dotnet - Unofficial typed C# client for the Jev System One API (Choice/Score/Noul). Project guide.
• jev-dotnet (CMaintz) - Unofficial zero-dependency .NET client for TypeSafe Jev Choice/Score/Noul (distinct from ukashanoor/jev-dotnet). Project guide.
• jev-droid (Android UI automation on a free emulator, with DeepSeek planning and Jev triaging screens) - Android UI automation agent (README in Portuguese) built on Google's free emulator, WHPX on Windows and local ADB. The agent reads the accessibility tree, DeepSeek V4.1 Flash picks one action per turn, and Jev 1.13 via OpenRouter triages screens and acts as a guardrail before actions run. Project guide.
• jev-drone - Explores Jev tactical judgments in a MuJoCo quadrotor simulation, with local flight control and a separate experimental tunnel policy. Project guide.
• jev-education (open experiments with Jev in education: a learning-path map that diagnoses a student's first mistake) - Public experiments using Jev in education without real learner data. The main demo, a learning-path map for first-year junior-high maths, takes an equation and the student's working, has Jev choose the most likely first mistake, then offers a check question, review material or the next problem, and holds off when probabilities are low; it shows API latency, estimated cost and the raw request and response. Japanese-language docs. Project guide.
• jev-effort - Unofficial Claude Code helper: TypeSafe Jev picks per-step reasoning effort and lease (OpenRouter/TypeSafe/Vercel). Project guide.
• jev-effort-router - Hermes plugin for Ollama:Cloud: TypeSafe Jev routes both model and reasoning effort per turn via llm_request rewrite. Project guide.
• jev-effort-router (hasanzaman22) - Claude Code plugin that asks Jev how much thinking each prompt needs and how risky it is, then runs the turn at the lowest effort level likely to get it right, leaning toward more effort when Jev is unsure. Project guide.
• jev-emoji - Dependency-free npm package that picks or ranks an emoji for a piece of text with a Jev Choice call, from a curated list of about 250, the full Unicode set or your own list, with a confidence fallback. Project guide.
• jev-enforce - Claude Code plugin: TypeSafe Jev checks every reply and file edit against CLAUDE.md / AGENTS.md rules and blocks broken rules in-turn. Project guide.
• jev-eval (calibrate Jev confidence thresholds on your labeled data) - Python CLI that runs your labeled rows through Jev once, then reports per-question accuracy, calibration (ECE and reliability tables) and risk-coverage, and recommends the confidence cut-points to gate on, written to a thresholds file the author's other Jev tools can load. Project guide.
• jev-eval-harness - Pre-registered research harness testing whether Jev's typed market judgments beat trivial baselines on hourly BTC/USD, with sealed holdout, purge, deflated Sharpe and bootstrap CIs; the measured answer was no. Project guide.
• jev-eyes - Local OCR/layout see() (and optional ask()) so TypeSafe Jev decides over inspectable image-derived text state. Project guide.
• jev-fabric - Single native executable (written in Bend) that owns an agent's long-running processes with deadlines and bounded logs, and asks Jev a typed choice/noul/score question only when explicitly told to. Project guide.
• jev-factcheck - Agent skill (English and Ukrainian) that asks Jev whether a verbatim source excerpt supports, contradicts or says nothing about a specific claim, keeping provenance. Project guide.
• jev-fanout-bench - Reproducible measurement that N TypeSafe Jev questions in one call bill shared state once (2,976 requests, raw JSONL). Project guide.
• jev-feels - Ruby gem for TypeSafe Jev as feels? / decide / score and Rails validations (distinct from BAML feelings and ruby_decision_model). Project guide.
• jev-filter (ByteBell) - Claude Code plugin: after grep, TypeSafe Jev noul-filters which files matter (~1% of LLM file-triage cost per upstream bench). Project guide.
• jev-flight-agent - Browser-use flight search demo: TypeSafe Jev picks DOM actions; small LLM types text only (Playwright/CDP). Project guide.
• jev-fn (call Jev as an ordinary Python function that returns a probability) - Small Python wrapper around the official TypeSafe SDK that exposes Jev as one callable: if jev("Is this customer angry?") > 0.9: flag_for_human_review(). Set one environment variable (TypeSafe or OpenRouter) and use it in if-statements, loops and pipelines; sync and async clients, model choice through configure(). Project guide.
• jev-for-all - Shared TypeSafe Jev decision contract for OpenCode, Claude Code, and Hermes (skill routing, tool subset, browser). Project guide.
• jev-foundation-models - Swift 6 bridge: TypeSafe Jev as an Apple Foundation Models LanguageModel for @Generable decisions (distinct from TypeSafe Swift client). Project guide.
• Jev-Frame - Python framework for typed, auditable Jev decisions in agents: describe judgments, candidate sets, completion contracts and budgets, and it compiles the Jev questions, tracks evidence with provenance and returns a supported typed result or an explicit unresolved outcome. Project guide.
• jev-frontier-bench - Reproducible typed-decision benchmark: TypeSafe Jev 1.13 vs Claude/GPT/Kimi/MiniMax/DeepSeek on 200 items (accuracy, calibration, cost). Project guide.
• jev-fuse - Governed execution/safety-gate/audit proxy for TypeSafe Jev and local Laya engines. Project guide.
• jev-gate (MongLong0214) - Claude Code plugin: TypeSafe Jev Gate/Router/Evidence plus Compact/Output (no LICENSE file; ≠ Neoo-Blue jev-gate). Project guide.
• jev-gate (Neoo-Blue) - Claude Code plugin: TypeSafe Jev checks the plan vs your request before edits, then reviews the final summary until gaps are closed. Project guide.
• jev-gates - Composable three-valued semantic logic circuits from TypeSafe Jev judgments and exact rules, with mock/offline demos (distinct from typesafe-agent-gates). Project guide.
• jev-gateway - Local LLM gateway that asks TypeSafe Jev which tool to call for Codex, Claude Code, OpenCode, or Gemini while other traffic uses your usual LLM. Project guide.
• jev-gateway (TexasOct) - OpenAI-compatible session-aware model-routing gateway with optional typed Choice decision providers (AGPL; distinct from vinilana/jev-gateway). Project guide.
• jev-gateway-bench - Benchmark suite for vinilana/jev-gateway: coding agents on chess-engine tasks with Jev routing on vs off. Project guide.
• jev-gen-detect (AI-generated image detector) - Rust web UI and CLI that asks Jev Image whether a picture is generated or authentic, or which generator made it, and asks Jev Box for the most telling artifact region, showing probabilities, latency and cost. Project guide.
• jev-git-graph - Evidence-backed TypeSafe Jev relationship graph for Git branch consolidation. Project guide.
• jev-governor - Claude Code mod for Jev-guided model/effort routing, verbatim context compaction, and output trimming. Project guide.
• jev-gpt (single-model GPT-6.1 Terra proxy for Codex: Jev picks reasoning effort and resumes tasks that stop early) - Local proxy between Codex (or any Chat Completions client) and one GPT model, GPT-6.1 Terra by default: before each generation Jev reads the turn so far and sets reasoning effort from low to xhigh, and when the model stops before the task is done Jev detects it and the proxy resumes. A sibling of the listed jev-sol-terra by the same author. Chinese-language docs. Project guide.
• jev-grievance-triage (auditable grievance triage for development finance) - Reads one grievance about a development project and returns a structured, auditable triage decision: Jev answers 23 narrow questions as calibrated probabilities in one request, versioned thresholds turn them into flags, and plain Python policy rules decide routing, urgency and human review. Project guide.
• jev-ground - Evidence-grounded factuality verification for LLM and agent outputs using a Jev provider. Project guide.
• jev-guard - Coding-agent security hooks: TypeSafe Jev risk-scores tool calls with session context (deny/ask/allow), flags prompt injection, and scans skills across Claude Code, Codex, Cursor, and more. Project guide.
• jev-guard (CMaintz) - Portable TypeSafe Jev allow/block/hold guard for agent tool calls with LangChain and Vercel AI SDK adapters (distinct from leepokai/jev-guard). Project guide.
• jev-guard (klauswg) - Exchange deposit/withdrawal risk triage: TypeSafe Jev judges; Java code adjudicates (distinct from leepokai/CMaintz jev-guard). Project guide.
• jev-guard (muratcakmak) - Claude Code hooks: TypeSafe Jev probability-scored rules deny bad edits/deploys; fail-open if scorer unavailable. Project guide.
• jev-guard (rudra72r) - Python LLM-app guardrails library: TypeSafe Jev (or local/Cloudflare backends) scores inputs/outputs with allow/review/block policies—distinct from leepokai/jev-guard agent hooks. Project guide.
• jev-guard (yelkhanyergali-sys) - PI Mono extension: in-flight terminal pruning (prompt-cache safe) and surgical diff guard using TypeSafe Jev—distinct from rudra72r/leepokai jev-guard variants. Project guide.
• jev-guardbench - Bench: can TypeSafe Jev / open-source Kev replace LLM-as-judge in agent guardrail callbacks? (license unspecified). Project guide.
• jev-guardrail-benchmark - Benchmark of TypeSafe Jev content-safety guardrails in WSO2 AI Gateway vs Azure Content Safety and an LLM judge. Project guide.
• jev-guardrails (deepansh-saxena) - Eval harness: same 25 guardrail rules on a mock support agent comparing LLM-as-judge JSON vs TypeSafe Jev (license unspecified). Project guide.
• Jev-guided CRANE+ imitation learning (Gazebo pick-and-place data generation where Jev selects the phase, then ACT training) - Course project that generates image-conditioned pick-and-place demos in Gazebo: Jev selects the high-level phase, MoveIt 2 plans the trajectory and ROS 2 controllers execute it; recordings become a LeRobot dataset for ACT training. The README reports limits honestly, including incomplete full-task results. Project guide.
• jev-harness - TypeScript decision harness: confidence-gated TypeSafe Jev policies, shadow mode, recipes, and jev-eval (distinct from super-jev / jev-layer). Project guide.
• jev-harness (Adityakk9031) - Tests a coding-agent setup: replays disguised dangerous commands against your hooks and permission rules, checks whether a gate's confidence is calibrated, probes instruction loss after compaction and compares AI reviewers, with Jev as an optional judge or gate. Project guide.
• jev-harness (Jev judgment gates for Claude Code) - Claude Code plugin with judgment gates: a Stop gate that catches turns ending by announcing work not yet done, and tool gates that keep secret files away from non-Anthropic providers, with rules first, Jev only where a rule cannot decide, and shadow mode before enforcement. Project guide.
• jev-harness (TypeSafeAI) - Research proposal-review harness: LLM proposes, Jev answers narrow questions, code leaves host evidence (TypeSafeAI community; distinct from AntonioCoppe/jev-harness). Project guide.
• jev-healthcare-lab - Open Jev vs DeepSeek comparison on 96 healthcare tasks across 12 scenarios (quality/latency/cost). Project guide.
• jev-herdr - Spawn Claude Code agents in Herdr with TypeSafe Jev choosing model and effort per agent. Project guide.
• jev-hook-guard - Standard-library Python module and example hook implementing the checks a Claude Code hook should run before sending session text to TypeSafe Jev or another HTTP judge: kill switch, exclusion words with session pins, refusal of secret dumps, masking of keys and personal data, length cut, leftover re-check, pending log line and an https-only, no-redirect POST. Project guide.
• jev-hooks - Claude Code commit-reviewer plugins driven by typed /v1/systemone decisions (self-hosted rizzo-flow or TypeSafe Jev) with a measured question bench. Project guide.
• jev-hooks by austpryb (nine Claude Code hooks backed by Jev) - Claude Code plugin with nine Jev-backed hooks: a Bash and an edit risk gate, compaction triage that keeps decisions, subagent report verification, loop detection, a stop self-check, subagent model routing, prompt routing and output narrowing. It fails open. Project guide.
• jev-humanizer - Anti-slop skill: TypeSafe Jev flags AI-sounding paragraphs so the writing model rewrites only those spans (RU-focused humanizer). Project guide.
• jev-imdb-benchmark - Benchmark TypeSafe Jev on IMDB reviews (sentiment/spoilers/quality) with published accuracy, latency, and cost. Project guide.
• jev-in-codex - Codex MCP tools for TypeSafe Jev-ranked capability selection, workspace search, and output triage (experimental MVP). Project guide.
• jev-inbox-triage - Store-inbox triage where TypeSafe Jev answers eight parallel questions per email (topic, urgency, sentiment and more) and rules route it to one of four lanes with every decision logged; the author reports 19/20 test emails in the right lane and no risky email automated. Project guide.
• jev-inference-router - Observable TypeSafe Jev decision routing with local Apple Silicon LLM inference, retrieval, and reproducible evaluations. Project guide.
• jev-intent-benchmark - Research check for indirect prompt injection at the moment an agent acts: asks Jev whether the act carries out something a tool result asked for but the user did not, with a benchmark script on InjecAgent. Project guide.
• jev-interceptor - Claude Code UserPromptSubmit hook (repo jev-setup) that asks TypeSafe Jev a few typed questions about every prompt and tells Claude where the work should run — on Claude or an open-model lane (ClinePass via pi, or Codex) — returning the answer as additionalContext; a one-command installer also sets up companion skills, a fast Jev compaction plugin and a spend log, with OpenRouter as a backup path. Project guide.
• jev-issue-radar - Local read-only dashboard: TypeSafe Jev (OpenRouter) classifies GitHub issue duplicates with evidence passages for maintainer review. Project guide.
• jev-java - Unofficial Java 17+ SDK for TypeSafe Jev with OpenRouter/Vercel adapters and optional Spring Boot starter (jev-typesafe 0.1.1). Project guide.
• jev-judge - Pytest/Vitest-style CI evaluations of LLM and RAG outputs: declarative YAML tests (faithfulness, relevance, hallucination, guardrails) answered by TypeSafe Jev noul/score questions instead of an LLM-as-a-judge, with JSON/Markdown reports and an MCP server. Project guide.
• jev-judge-mcp (PyModel) - Typed judgment MCP tools over TypeSafe Jev; policy decides auto/review/escalate (distinct from gecm0). Project guide.
• jev-judge-rank - Three standard-library Python scripts (and an agent skill) that use TypeSafe Jev to rank a list on several written criteria, merge near-duplicates under one representative, and recall the corpus entries that help with a one-sentence situation; every request and response is logged locally, with a recorded example run and its costs. Project guide.
• jev-lab (gregoryjayuk) - Two standard-library Python experiments with Jev: a feed sieve that judges a week of posts from 500+ RSS, Atom, and YouTube feeds against a reader profile and writes a reading list, and a bookmark re-filing eval of agreement, calibration, and cost. Project guide.
• jev-lanes (confidence-based routing and escalation for Jev in LangGraph) - Python library that routes each Jev classification into lanes by confidence: accept, retry with a fallback LLM, or pause for a resumable human review in a LangGraph StateGraph, with a serializable trace of every route. Project guide.
• jev-layer - Portable System-1 decision layer for agent harnesses: host-owned capability routing, receipts/replay, optional supervision, demo/OpenRouter/TypeSafe providers. Project guide.
• Jev-LCT - Open System One decision engine (Looped Calibration Transformer) with Jev-compatible Choice/Noul/Score serving and HF Qwen-based checkpoints—independent of hosted Jev. Project guide.
• jev-lgtm - GitHub Action that asks TypeSafe Jev to approve a PR (LGTM) or flag Review Advised from the diff alone. Project guide.
• jev-libero - Fine-grained LIBERO robot control study loop: TypeSafe Jev layered decisions with local physics previews and inspectable records. Project guide.
• jev-linkedin-saved-classifier - LinkedIn saved-posts board classified into filters by TypeSafe Jev. Project guide.
• jev-lint - Ast-grep + TypeSafe Jev natural-language semantic linter (name/body drift, stale comments, weak tests); distinct from huntedman/JevLint. Project guide.
• jev-lint (ckorhonen) - Fuzzy agent linter (jevlint.dev): hooks check agent-written code against team best-practice packs while the agent works (distinct from mizchi/jev-lint and huntedman/JevLint). Project guide.
• jev-linter-action - GitHub Action: configurable TypeSafe Jev yes/no suites over repo files with probability thresholds (distinct from jev-ci-selector and jev-lint CLIs). Project guide.
• Jev-Lite - Open System One–style FastAPI decision engine (Qwen backbone) inspired by TypeSafe Jev; independent of hosted Jev. Project guide.
• jev-lite by Nefelim584 (395M ModernBERT typed decision model benchmarked against the Jev API) - Small typed decision model built to replace an LLM-as-a-judge on narrow checks: given a state and bool, choice or score questions, it returns calibrated probabilities and a confidence score from a ModernBERT-large encoder with three heads. Scripts run the same sampled test records through Jev's API for paired accuracy, speed and repeatability comparisons with NLI and Laya baselines. Distinct from Atulmishra22/jev-lite. Project guide.
• jev-llm-guard - Contextual OWASP LLM Top 10 guardrail powered by TypeSafe Jev, with a Turkish interactive demo. Project guide.
• jev-llm-router (Pi) - Pi extension that uses Jev to select a model and thinking level for each user task, keeps the pair through tool continuations, and supports subagents selecting jev/auto. Project guide.
• jev-local - Local CPU-only Jev-style 0.8B decision service plus an oh-my-pi extension that points the agent's typesafe judge provider at it, reading yes/no logits from llama.cpp. Project guide.
• jev-loop (King4s) - TypeSafe Jev decides, Claude Code or Hermes builds—MCP server + skill (distinct from lvzhaobo/jev-loop). Project guide.
• jev-macos-loop - Native Apple silicon macOS computer-use agent: local OmniParser/Vision/Accessibility perception with text-only Jev action choice (BYOK Vercel/OpenRouter/TypesafeAI). Project guide.
• jev-mail-triage (Gmail and Fastmail filing with Jev) - Files each new Gmail, Google Workspace or Fastmail message with one Jev-chosen label (political asks, newsletters, receipts, junk and more), archives or spams only above set thresholds, and ships a small web console for editing rules and correcting mistakes. Project guide.
• jev-map - Maps each Python function to the tests that run it for coding agents; optional TypeSafe Jev scores candidate links that static analysis misses, checked against real test runs. Project guide.
• jev-max - Jev browser agent aimed at pages ordinary Jev agents miss: a piercing snapshot walks open shadow roots and same-origin iframes into one indexed element table, a vision fallback handles canvas, clicks are occlusion-checked, one Jev request per cycle picks the operation and target (TypeSafe or OpenRouter), a small LLM writes text only for fill/select, and an independent Jev check verifies DONE; ships as a Python library, CLI and MCP server. Project guide.
• jev-mcp - MCP server with ten purpose-built TypeSafe Jev judgment tools (verify, screen, find, classify, review, gate, and more). Project guide.
• jev-mcp (Afloat16) - Unofficial conservative MCP server for TypeSafe AI Jev (stdio) so agents can ask typed System One questions. Project guide.
• jev-mcp (drmedia) - General-purpose MCP server for System One decision models: MCP clients ask typed yes/no, choice and score questions about structured state and get calibrated probabilities from TypeSafe Jev, OpenRouter (Jev, Cloudflare Clef) or a local server, with batching, retries, image input notes, stdio/HTTP transports and Docker. Distinct from the separately listed jev-mcp. Project guide.
• jev-mcp (legostin) - MCP server: Claude/Codex agents drive real Chrome with TypeSafe Jev (~100 ms typed decisions, confidence gates, traces); distinct from jkudish/jev-mcp. Project guide.
• jev-mcp (pyck-ai) - Fourteen fail-closed MCP tools over Jev via OpenRouter System One (jev_score, jev_gate, jev_ask, …)—distinct from markylaredo/openjev-mcp. Project guide.
• jev-mcp (rules over OpenRouter) - Personal MCP server that lets Claude and other assistants ask Jev yes/no and multiple-choice rules written in a YAML file, returning a decision with a probability and flagging uncertain results for the user. Project guide.
• jev-mcp-router (ini8labs) - MCP gateway where TypeSafe Jev answers yes/no per server and tool, so the agent LLM only sees the tools a request needs. Project guide.
• jev-mcp-server - Python MCP server for official TypeSafe Jev choice/score/noul plus compare/verify/classify and a one-command client installer. Project guide.
• Jev-Mem - System-One–controlled agentic memory: TypeSafe Jev steers admission, multi-relational linking, and adaptive retrieval before a text model answers. Project guide.
• jev-mem (must-lioncho) - Memory plugin for Claude Code, Codex, and Antigravity that keeps claude-mem style fast search as a first pass and lets Jev judge which past records are relevant to your prompt, so one good record goes in or none. Project guide.
• jev-mnist - Shows all 10,000 MNIST test digits to Jev 1.14's image input as a 0 to 9 Choice question: 92.45% accuracy at 0.17 s per image, with a confusion matrix that shows a strong bias toward answering 7. Project guide.
• jev-mobile-pilot (MCP server that lets coding agents drive iOS Simulators and Android devices, with Jev picking each step) - MCP server and CLI for coding agents: code reads the iOS Simulator or Android accessibility tree, Jev picks the next element and decides whether the goal is done, and your agent only plans. Screenshots are never sent to Jev; tests run against a fake device and a fake Jev. Project guide.
• Jev-Mod - Open-source Discord moderation bot where TypeSafe Jev rules detect spam, phishing, harassment, hate, threats and explicit text, with log/delete/timeout actions, local phrase and mention filters, exceptions and a case dashboard; use the hosted service with your own TypeSafe key or self-host. Project guide.
• jev-mod (VictorGambarini) - Claude Code mod that hands small decisions to a decision model — Jev via TypeSafe, OpenRouter, Venice or OpenCode Zen (or your own server): per-turn routing lanes that set effort and model, the one installed skill a prompt needs, screening of AI-directed instructions after WebFetch / WebSearch / MCP / fetching Bash, and /compact-jev, a summary-free compaction keeping only turns marked keep. Every decision fails open. Distinct from the listed Discord bot Jev-Mod (undeemed). Project guide.
• jev-model-router (m-esm) - Claude Code hook where TypeSafe Jev scores each prompt and picks model and effort (distinct from the claude-code-templates mod). Project guide.
• jev-navigator - Python code-navigation library and jvn CLI for agents: a model-free index (definitions, callers, references, imports, git history) plus small typed Jev judgments — check, pick, score — composed into searches like “find the code this sentence describes”. Project guide.
• jev-netlify-mcp - Two parts for using TypeSafe Jev from Claude Code, Codex CLI, Claude Desktop or any MCP client: a key-locked Netlify function that exposes POST /v1/systemone and GET /v1/models through Netlify AI Gateway (which injects the gateway credentials, so no TypeSafe key is needed), and an MCP server with jev_evaluate / jev_models tools plus question-writing instructions; it also works directly against TypeSafe with your own key. Project guide.
• jev-oas-sentinel - Compares OpenAPI documents with deterministic structural checks plus TypeSafe Jev semantic questions; policy code owns pass/review/block. Project guide.
• Jev-Omni - Open multimodal System One–style decision classifier (text/image/audio/video → option probabilities) fine-tuned from Gemma 4 12B IT; independent of official Jev. Project guide.
• jev-omni.js - Run independent Jev-Omni decision classifier in-browser on WebGPU (onnxruntime-web; text+images). Project guide.
• jev-ood-calibration - Independent TypeSafe Jev calibration study with published raw responses: three public benches plus 900 OOD synthetic tickets. Project guide.
• jev-open-calibration (Jev's calibrated-decision interface rebuilt on scikit-learn and tested against Jev) - Code, data and figures for the Towards Data Science article "Jev Is API-Only. Its Best Idea Is Open.": rebuilds Jev's typed Choice, Score and Noul interface on scikit-learn with temperature scaling, then compares calibration, cascades and a none-of-the-above trap with live Jev. Project guide.
• jev-opus - Claude Opus 5.5 with TypeSafe Jev re-picking effort every step via per-message statements (prompt-cache friendly). Project guide.
• Jev-Orchestrated Decision Mesh (CIDM) - Training-free orchestration design where Jev chooses the next unit and reviews each output across a five-unit network, with a Codex CLI broker and skill. Project guide.
• jev-orchestrator (milan6190) - Workflow orchestration for System One models: batch, gate, trace, and calibrate thousands of typed Jev decisions. Project guide.
• jev-packs - Evidence-gated Jev question-pack registry with golden cases and a reproducible offline multi-backend scoreboard. Project guide.
• jev-permission-gate (madisonrickert) - Claude Code auto-mode tool-call gate using TypeSafe Jev to allow, deny, or defer to the built-in classifier. Project guide.
• jev-phone - Phone agent: TypeSafe Jev picks indexed UI actions; phone-use executes on iOS Simulator, Android, or cloud phones (BYOK via AI Gateway). Project guide.
• jev-php (f-lombardo) - PHP client library for TypeSafe Jev / System One APIs (LGPL-2.1). Project guide.
• jev-pi (weiping) - pi package port of weiping/jev-claude-code: TypeSafe Jev permission gate, output ladder, conditional context, and agent router (extension + skills + prompts). Project guide.
• jev-pi-token-reduction - Pi coding-agent extension: TypeSafe Jev trims tool-output chunks before the model sees them (~15% cost on held-out reads). Project guide.
• jev-pii (PII and privacy-regime discovery in DuckDB SQL with typed Jev answers) - DuckDB functions that return Jev's typed answers as SQL columns: per row (jev_pii_row, jev_ask) or per column from a sample (jev_pii_table) to find personal data and the privacy regimes that apply, feeding a keep, mask, hash, drop or review policy written in code; also runs in the browser through Pyodide. Project guide.
• jev-pii-checker - CLI that finds PII spans and sensitivity with TypeSafe Jev plus regex/segmentation layers (scanned text leaves the host). Project guide.
• jev-pilot (Akramovic1) - Claude Code plugin: TypeSafe Jev picks effort, subagent model, strategy, and one skill per prompt (distinct from JevPilot driving demo). Project guide.
• jev-pilot (wangzhezbz) - All-in-one Codex plugin: TypeSafe Jev for reasoning-effort routing, context filtering, and workflow assistance on macOS/Windows/Linux (distinct from Akramovic1/jev-pilot). Project guide.
• jev-plan-review (Claude Code plugin that checks a plan keeps every requirement you stated) - Claude Code plugin that collects what you said (prompts, answers, rejected-plan feedback, pasted text), splits it into statements and has Jev check each one against the plan, flagging anything missing or contradicted with a confidence score; it runs as a plan-mode hook or a review command and also lists decisions you never asked for. Project guide.
• jev-playground - OpenRouter workbench for TypeSafe Jev: typed noul/choice/score questions with calibrated probabilities, latency/tokens/cost metrics, and Postman/cURL export. Project guide.
• jev-playground by exponen-agi (16 interactive simulations and runnable cookbooks for designing with Jev) - Engineer's playground for Jev: sixteen browser simulations (no install, no key, nothing called) showing how a fast, calibrated, typed decision model changes an architecture, plus runnable Python cookbooks for triage cascades, RAG gatekeeping, agent guardrails, output verification, lead scoring and model routing, and a plain-English guide with a where-it-breaks section. Project guide.
• jev-plays - TypeSafe Jev plays Craftax: macro/raw action Choice with optional LLM planner-as-facts and a local web UI. Project guide.
• jev-plays-runescape - Archived hobby demo where Jev chooses woodcutting and banking actions on a local 2004Scape server, with live probabilities and local safety vetoes. Project guide.
• jev-playwright (arthurfiorette) - Playwright reporter: TypeSafe Jev selects which E2E tests to run from the diff (distinct owner). Project guide.
• jev-poke (Jev plays Pokémon Showdown) - Node.js bot (README in Chinese) that plays Pokémon Showdown VGC doubles: at each team preview, turn and forced switch it sends structured questions about the full battle state to Jev through OpenRouter's Decisions API, with local fallbacks and an optional advisor model. Project guide.
• jev-polymarket-test - Study of whether a TypeSafe Jev Noul beats Polymarket prices on resolved markets (blind vs priced, memorisation check). Project guide.
• jev-pr-judge - Typed pull-request verdicts via one parallel TypeSafe Jev call, code-owned policy profiles, Next.js UI, and a sticky-comment GitHub Action. Project guide.
• jev-pr-profiler - GitHub Action: TypeSafe Jev PR risk profile + review-depth outputs for CI (never merges alone). Project guide.
• jev-pr-quality - GitHub Action for TypeSafe Jev-assisted pull request quality reviews with a live multi-repository RawTree dashboard. Project guide.
• jev-pref - Turns AGENTS.md / project preferences into a TypeSafe Jev semantic linter for coding-agent diffs (npx jev-pref). Project guide.
• jev-preflight - Claude Code Stop-hook risk check: TypeSafe Jev scores eight axes over a redacted turn diff, with optional one-shot reinspection. Project guide.
• jev-project-context - Evidence-first experiment memory skill for coding agents with optional TypeSafe Jev semantic triage. Project guide.
• jev-prompt-sentry - Anthropic Messages reverse proxy that screens jailbreaks/injections with one batched TypeSafe Jev call (PolyForm Noncommercial). Project guide.
• jev-proxy - Local TypeSafe System One proxy that records, caches, and replays calls with a web inspector. Project guide.
• jev-pruner - Prunes eligible Bash stdout with Jev before it reaches Claude Code or an opt-in Codex wrapper, while archiving the original output locally. Project guide.
• jev-py-sdk - Unofficial Python SDK and CLI for typed Jev noul/choice/score judgmen