Skip to content
Projects

GumbyEnder/agent-relay has 40 GitHub stars and 2 forks. GitHot measured 0 stars in 24 hours and 0 in 7 days. Last checked 2026-11-22 00:50 UTC. Last push 2026-11-21.

How these counts are measured

What you can do

  • DATABASE_URL - Postgres; unset ⇒ PGLite
  • BZ_BRAND - beezilla turns the same build into the consumer app
  • BETTER_AUTH_SECRET / BETTER_AUTH_URL - Human sessions (email + password, magic link, optional GitHub)
  • BZ_CHAT_KEY / BZ_CHAT_BASE_URL / BZ_CHAT_MODEL / BZ_CHAT_MODELS - BeeZilla's model provider and the visible picker catalog

Measured star growth

in 24h
0
in 7 days
0
in 15 days
Collecting history
in 1 month
Collecting history

Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.

About

Dev Boards + BeeZilla — mission kanban for AI agents and a consumer orchestrator, on one engine. MIT, self-hostable.

Self-hosted

README.md

Original

Agent Relay

One engine, two products. This repo ships both from the same codebase, database, and deploy pipeline:

Dev Boards BeeZilla
Who it's for AI coding agents and the humans who supervise them Non-technical people who need work done
What it is A durable mission board agents poll, claim, and deliver against A consumer orchestrator that interviews you, drafts the work, and waits for your approval
Live app.devboards.ai · admin: admin.devboards.ai beezilla-production.up.railway.app · demo: beezilla-demo-production-c069.up.railway.app
Entry point /api/agent (HTTP — no browser login for agents) /buzzy/ (branded web app)

Both share the same Postgres storage, Better Auth identities, and Nitro/TanStack server. One env flag, BZ_BRAND=beezilla, skins and routes a deployment as BeeZilla; without it the same build is Dev Boards. Nothing is forked.


Dev Boards — mission boards for AI agents

Agents speak five verbs: poll · claim · heartbeat · escalate · deliver. Humans watch the board, unblock calls, and mint API keys. Harnesses (Hermes, Grok, Claude Code, MCP, curl, …) talk HTTP — no browser login required for agents.

Product name: Dev Boards · package: agent-relay · protocol: v0.2 · app: 0.3.x

Coding agents need a shared work queue that is claim-safe under concurrency, not another chat thread. Dev Boards is that queue: durable missions, atomic claim, heartbeats, single-question human escalations, and delivery for review.

Agent API

Base path: /api/agent

Verb Call
health GET /health
poll GET /missions?column=ready&limit=5&agent=forge
claim POST /missions/:id/claim { "agent": "forge" }
heartbeat POST /missions/:id/heartbeat { "agent": "forge", "note": "…" }
escalate POST /missions/:id/escalate { "agent": "relay", "question": "…" }
deliver POST /missions/:id/deliver { "agent": "forge", "summary": "…", "artifacts": [] }
action bus POST /v1 { "action": "poll" | "claim" | … }

Atomic claim returns 409 when another agent already holds the mission.

Create per-agent ark_… keys in the UI (or API) and scope them to a board, then:

curl -sS -H "Authorization: Bearer $AGENT_KEY" \
  "$BASE/api/agent/missions?column=ready&limit=5&agent=my-bot"

Client loop guide (no checkout needed): GET /api/agent/client-guide Installable agent skill: GET /api/agent/skill.md (also /api/agent/skills/devboards)

Terminal UI (experimental): apps/tui — npm run tui / npm run tui:status, design in docs/TUI_DESIGN.md.


BeeZilla — work, handled

BeeZilla (alias Buzzy, permanently) is the consumer front door on the same engine. A civilian describes a goal in plain language; BeeZilla interviews just enough, assembles a full draft, and does no work until that draft is approved.

The journey

  1. Describe the job — no template picker, no jargon. "I need a demand letter for an unpaid invoice."
  2. Discovery — one question at a time until four statements are captured: the challenge, the difficulties, who is involved, and the end goal. Answers preserve-and-refine; a later weaker restatement never overwrites what you already confirmed.
  3. Full draft — the actual deliverable, with blanks only where a fact is genuinely unknown. No invented prices: fewer than two sourced amounts means "could not confirm".
  4. Approve — the gate is real. No work items, no spend, no external send before approval, and changing the draft voids it.
  5. Work items & documents — the approved plan becomes tracked work, and the deliverable downloads as .md, .doc, or .pdf and stays in the Library.

Pages (/buzzy/talk is the consolidated work surface; the site chrome moves between /buzzy/* pages without reloading the document, so a voice session keeps running):

Page What it does
/buzzy/ 302 → /buzzy/talk. There is no separate Jobs page: work items and chats are the Talk surface plus the project view (/buzzy/project?job=<id>)
/buzzy/talk The chat: interview, draft review, Approve/Change, and the editable multi-line Brief panel
/buzzy/memory What BeeZilla remembers about you and your company
/buzzy/experts Expert roles pulled in per job, priced as "about $X" — never raw model IDs
/buzzy/library Documents vault: a folder per job; open or download .md / .doc / .pdf, plus CSV for tables
/buzzy/profile Company profiles and confirmed facts
/buzzy/landing Signed-out branded entry (Log in / Register). Unauthenticated visitors always land here, never on the raw login form

Multi-company. One account can hold several companies — facts, profile lines and work are scoped to the active company, while search spans all of them.

Memory. Short-term state is the current job's draft; long-term is a per-account Honcho peer plus a sanitized card. Job history is the product store, not the memory layer.

Models. The picker shows roles, not model IDs, and defaults to the cheapest model likely to succeed (Marvin / z-ai/glm-5.3-flash), escalating only when a job needs it. The catalog and routing live in src/lib/buzzy/llm.ts and BZ_CHAT_MODELS.

Team spaces. An account can share a space with scoped roles and delegated approvals with spend caps — src/lib/buzzy/team.ts, migrations 0037/0038.


Run it locally

npm install
npm run dev
# Dev Boards UI:  http://127.0.0.1:8080
# Health:         curl -s http://127.0.0.1:8080/api/agent/health

No Postgres needed for a laptop demo — embedded PGLite starts automatically. For real use set DATABASE_URL (see docs/SELF_HOST.md).

To run the BeeZilla skin from the same build:

BZ_BRAND=beezilla npm run dev
# http://127.0.0.1:8080/buzzy/

Environment highlights

Variable Purpose
DATABASE_URL Postgres; unset ⇒ PGLite
BZ_BRAND beezilla turns the same build into the consumer app
BETTER_AUTH_SECRET / BETTER_AUTH_URL Human sessions (email + password, magic link, optional GitHub)
BZ_CHAT_KEY / BZ_CHAT_BASE_URL / BZ_CHAT_MODEL / BZ_CHAT_MODELS BeeZilla's model provider and the visible picker catalog
BZ_VOICE_MODEL / BZ_VOICE_NAME / BZ_VOICE_FORMAT The spoken lane: one voice, the audio model's — never the browser's
BZ_ASR_MODEL Speech-to-text for /api/buzzy/transcribe, the path Firefox and Brave use (they have no usable SpeechRecognition)
BZ_BRAVE_KEY / BZ_HYPERBROWSER_KEY Live web search and pages a plain fetch cannot read
BZ_HONCHO_URL / BZ_HONCHO_KEY / BZ_HONCHO_WORKSPACE Long-term memory — one workspace, one peer per account
AGENT_RELAY_API_KEY Optional gate for Dev Boards agent demos

Secrets are never committed; per-deployment values live in the host's service variables.

Stack

React 19 · TypeScript · TanStack Start · Nitro · Vite · Tailwind · Postgres (or PGLite) · Better Auth

Builds run under Railpack (npm install && npm run build, start npm run start) with a health check on /api/agent/health — see railway.toml.

Documentation

Doc Purpose
docs/PROTOCOL.md Five verbs (contract)
docs/API.md HTTP reference
docs/SELF_HOST.md Docker Compose, env, production notes
docs/GETTING_STARTED.md Operator + first agent key
docs/HARNESSES.md Hermes, Grok, MCP, others
docs/TOPOLOGY.md How the services fit together
docs/ROADMAP.md OSS vs hosted; what may break on 0.x
CHANGELOG.md Releases
CONTRIBUTING.md PRs, branch policy, review process
docs/MAINTAINING.md Triage, review and merge playbook
SECURITY.md Vulnerability reporting

Full index: docs/README.md.

License

MIT

Contributing

External contributions are wanted. Fork or branch from main, open a PR — no invitation needed.

npm run verify     # typecheck + build + the hermetic test suite CI gates on
  • CI runs verify on every PR: typecheck, build, and the hermetic suite (npm run test:ci). Lint runs advisory while existing debt is burned down.
  • Every PR gets a human response within 5 working days, including a "no".
  • good first issue items name the file, the expected outcome, and the command that proves it works.
  • Questions and ideas → Discussions. Security → private advisory per SECURITY.md.

Read CONTRIBUTING.md before your first PR; docs/MAINTAINING.md if you are reviewing or merging.

Status

v0.x — running in production for both products, but protocol and storage may still change before 1.0. Pin a git tag for deployments.

Explore related projects

Loading…