Skip to content
Projects

gutomec/nirvana-os-engine

View on GitHub

gutomec/nirvana-os-engine has 83 GitHub stars and 24 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-15 05:16 UTC. Last push 2026-11-13.

How these counts are measured

Measured star growth

in 24h
0
in 7 days
Collecting history
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

Nirvana-OS engine: ready-to-run agentic operations. One sentence in, finished work out. Source-available engine for companies, squads and mind-clones. Runtime-agnostic: Claude Code, Codex, Gemini-CLI, Antigravity, Hermes.

Productivity

README.md

Original
Nirvana-OS: one sentence in, finished work out. Parallel agent workflows converge through quality gates.

Nirvana-OS

Ready-to-run agentic operations. One sentence in. Finished work out.

GitHub stars version CI license

macOS, Linux and WSL

curl -fsSL https://gutomec.github.io/nirvana-os-engine/install.sh | bash

Windows (PowerShell)

irm https://gutomec.github.io/nirvana-os-engine/install.ps1 | iex

Documentation · Packs · Illustrated install · Beta · Changelog

Read this in your language: English · Português · Español · 中文 · हिन्दी · العربية


You talk to your agent. Your agent runs Nirvana.

You already work with a terminal agent: Claude Code, Codex, Antigravity, Pi, Grok or another. One prompt gets one good answer. Real work takes a researcher, a writer, a reviewer and an operator, with a record of who did what.

Nirvana-OS gives your agent that organization. You ask in plain language, and the agent uses Nirvana to pick the right businesses, squads and mind-clones, dispatch them, check what comes back and show you the receipt. You do not learn a new interface and you do not type commands day to day. The commands further down are the reference your agent uses, and they are there for you when you want them.

What it is

Nirvana-OS is a Bun-native, runtime-agnostic multi-agent system. It runs on top of the agent you already use and never replaces it. It works with three kinds of content, all of them folders on your machine:

Pillar What it is Where it lives
Businesses Organizations with an org chart. A dispatched business runs as one agent that plays its seats, uses squads through their cards and channels mind-clones ~/businesses/
Squads Portable agent teams with agents, tasks and workflows the dispatched agent follows as its method ~/squads/
Mind-clones Persona DNA in five layers, so an agent thinks and writes with a specific expert's method ~/businesses/_library/dna/

The engine ships with no content. You create businesses, squads and clones by asking your agent, or you install a pack from squads.sh.

Install

This README describes 0.15.1, the current stable release, which the one-liner at the top of this page installs. Node is not needed. If Bun is missing, either installer puts it in your user folder (~/.bun) without sudo or admin rights and carries on.

Before you run it:

  • macOS: nothing to prepare.
  • Windows (PowerShell): nothing to prepare, and no administrator rights.
  • Linux (Ubuntu): sudo apt update && sudo apt install -y curl unzip.
  • Windows with WSL (Ubuntu): run wsl --install in a PowerShell opened as administrator, reboot, open the Ubuntu app and create your Linux user. Then install curl and unzip as on Linux and run the bash command inside Ubuntu. Your AI agent (Claude Code, Codex, Antigravity) must be installed inside Ubuntu too, and your projects belong in ~, not under /mnt/c, which is slow.

What it changes on your machine:

  • the engine in ~/.nirvana/skills;
  • the nirvana skill linked into every agent runtime it finds: ~/.claude/skills, ~/.codex/skills, ~/.gemini/config/skills (Antigravity), ~/.pi/agent/skills and ~/.agents/skills (OpenClaw);
  • nrv and nrv-hermes in ~/.local/bin, plus one PATH line in your shell profile (the user PATH on Windows);
  • audit hooks in Claude Code, Antigravity and Codex;
  • the empty content folders ~/squads, ~/businesses and ~/businesses/_library/dna.

Running it again is safe. Close the terminal and open a new one so nrv is on your PATH, and check with nrv --version; it prints 0.15.1. Then ask your agent: "Check that Nirvana-OS is installed correctly." It runs nrv doctor. Exit 1 means warnings only; exit 2 lists what failed.

Through curl | bash the installer asks no questions. To pick a performance profile at install time, pass it through: replace the final bash with bash -s -- --profile=balanced. You can also set it later by asking your agent (see Settings).

The step-by-step guide for each system is on the documentation site (in Portuguese), with a page for betas, and in INSTALL.md, with the other routes (the nirvana skill through skills.sh, installing from a clone).

Start a project

Open the folder you want to work in, start your agent there, and say: "Set up this folder as a Nirvana-OS project." The agent runs nrv init ., which writes .nirvana/project.yaml and a short note in AGENTS.md, CLAUDE.md and GEMINI.md. Deliverables land in outputs/ inside the project.

That note keeps your agent working as usual. It uses Nirvana only when your request calls for it:

  • you name it ("use Nirvana to…", "via nirvana");
  • you name or point to one of your businesses, squads or mind-clones ("my launch company", "the copy squad", "in Hormozi's voice"), including a pack name;
  • you ask for the work, or part of it, to run on another runtime ("use Codex to review it").

If you want Nirvana to orchestrate every deliverable in a project, say "orchestrate everything in this project with Nirvana". The agent runs nrv init . --orchestrators=always. Open a new session for the change to take hold. A project set up on 0.14 keeps its full contract until you run nrv init . again, which switches it to the on-demand note in place (--orchestrators=always keeps the full contract) and keeps your own lines.

From then on you talk to your agent in plain language and it uses Nirvana-OS. You do not type commands day to day.

What to ask your agent

You say Your agent runs
"What businesses / squads / mind-clones do I have?" nrv list-businesses, nrv list-squads, nrv list-clones
"Use Nirvana to write the launch copy for X." the harness: picks the business, writes a brief, nrv dispatch … --exec
"Run this through the copy squad, on Codex." nrv dispatch --squad <slug> … --exec --runtime codex
"Stop that run." nrv run-track stop <run>
"Change the headline in what you delivered." nrv revise <project> "<change>"
"Show me what that run did." nrv audit-view <project>
"Install the dependencies of all my squads." nrv activate --all
"Update Nirvana-OS." nrv update
"Update my Genesis Circle pack." nrv update genesis-circle
"Put Nirvana in economy mode." nrv config set execution.profile economy --global

Bigger requests work the same way:

Use Nirvana-OS to create a company called podcast-empire that produces, publishes,
and monetizes 3 podcasts at once. Each show has its own niche, an AI host, an
editorial calendar, and an independent monetization funnel. Around 7 employees.
Use Nirvana-OS to turn the public work of <author> into a complete AI mind-clone.
Use Nirvana-OS to produce a launch package: market research, landing-page copy,
and a competitive teardown.

The agent orchestrates and never writes the deliverable itself. Each business it picks gets its own brief and its own run, in parallel when the parts are independent.

One prose order in, a whole AI conglomerate assembles and delivers

Packs

Packs are curated businesses, squads and mind-clones sold on squads.sh. The flagship is Genesis Circle.

  1. Buy the pack on squads.sh and download its zip from your dashboard.

  2. Unzip the whole archive and open a terminal in the folder that holds setup.sh:

    • macOS: double-click the zip, type cd in Terminal, drag the folder onto the window and press Enter;
    • Windows: right-click the zip, choose Extract All, open the folder, right-click an empty area and choose Open in Terminal;
    • Linux and WSL: cd ~/Downloads && unzip <file>.zip -d <folder> && cd <folder>. On WSL, first copy the zip from /mnt/c/Users/<user>/Downloads to ~/Downloads.
  3. Run the setup:

    • macOS, Linux and WSL: bash setup.sh
    • Windows: powershell -ExecutionPolicy Bypass -File setup.ps1

    Setup installs Bun if needed, installs or updates the engine, copies the content into ~/squads, ~/businesses and ~/businesses/_library/dna, rebuilds the indexes and stores the pack's license in ~/.nirvana-license/packs/<pack>/. It never replaces the engine with an older release, so a beta stays a beta.

  4. Talk to your agent: "Use Nirvana to list what the Genesis Circle pack gave me", then ask for work by naming a business, a squad or the pack.

Activating squads. A squad's dependencies install on demand the first time it runs, in its own environment under ~/.nirvana/envs/<squad>. To prepare them all up front, ask your agent: "Activate all my Nirvana squads." It runs nrv activate --all, which covers Node and Python packages, system programs, API keys and environment variables, MCP servers and models. The agent gets the list of failures with the exact command that fixes each one, and asks your permission before anything that uses brew, sudo or winget, and for any API key. A first run on a big pack takes several minutes.

Updating a pack. Ask your agent to update it, or run nrv update <pack-slug>, for example nrv update genesis-circle. Components you edited are backed up under ~/.nirvana/backups/packs/<pack>/ before the update replaces them. If the update fails with a license error, download the pack again from squads.sh and run its setup once more; that also refreshes the license.

Updating the engine

Ask your agent to update Nirvana-OS, or run nrv update. It backs up ~/.nirvana/skills first and keeps your settings and content. nrv update --check names the release it would install without changing anything.

An update never moves the engine to an older release. A stable engine moves only to a newer stable, so a 0.14 engine moves to 0.15.1. A beta moves to the newest greater release, beta or stable, so a beta becomes the stable release that follows it. After an engine update, nrv update also checks every installed pack that has a license and says when one has a newer version.

Betas

Betas are GitHub prereleases, and the one-liner at the top never installs one. A beta is installed by pinning its tarball: replace v<version> below with its tag from the release list. The tag has a leading v; without it the download returns 404.

macOS, Linux and WSL:

curl -fsSL https://gutomec.github.io/nirvana-os-engine/install.sh | NIRVANA_ENGINE_URL=https://github.com/gutomec/nirvana-os-engine/releases/download/v<version>/nirvana-os-engine.tar.gz bash

Windows (PowerShell), as one line:

$env:NIRVANA_ENGINE_URL='https://github.com/gutomec/nirvana-os-engine/releases/download/v<version>/nirvana-os-engine.tar.gz'; irm https://gutomec.github.io/nirvana-os-engine/install.ps1 | iex

Then, if the window is still open, run Remove-Item Env:NIRVANA_ENGINE_URL: while the variable is set, nrv update reinstalls the pinned tarball. The Windows commands use single quotes because a double quote lost in a paste leaves PowerShell waiting on >>.

Once on a beta, nrv update and every pack setup keep you on the beta channel: newer betas, then the stable release that follows.

Back to stable. On Windows, open a new PowerShell and run Remove-Item Env:NIRVANA_ENGINE_URL; an error saying it does not exist is fine. Run the plain one-liner from the top of this page. It installs the latest stable over the beta and keeps your content, packs and settings. Check that nrv --version no longer shows beta, then run nrv init . again in each project you used with the beta (details).

Settings

Settings change when you ask. "Put Nirvana in economy mode", "maximum quality" or "what is configured?" are enough; your agent runs nrv config and tells you what changed.

A profile moves several settings at once:

Profile Effort Context ceiling per worker Review Visual checks by workers
max xhigh none (runtime default) every delivery, up to two rounds on
balanced high 400k tokens when a rule asks for it off
economy medium 200k tokens only on request off

A profile never picks the model; your runtime's own configuration does. A setting you change explicitly wins over the profile. nrv config list shows every setting, its value and where it came from. The full reference is docs/configuration.md. Profiles exist only from 0.15 on.

Per project. The orchestrator writes each brief with a brief template. The engine ships four: slim (the default: short labelled lines, the user's words kept in a file), standard (the six sections), detailed (the six plus a suggested structure, the method and one criterion per item) and karpathy, after Andrej Karpathy's post on X and his 2025 Year in Review: slim plus a plain-language summary, a diagram and, when it helps, an HTML explainer, at more tokens per delivery. nrv config set briefing.template detailed --project picks one for a project, and nrv brief templates lists them. To change one or add your own, say juridico, run nrv brief init juridico --from standard --project: the copy in .nirvana/brief-templates/ wins over ~/.nirvana/brief-templates/ and over the engine's. briefing.altitude still works as a deprecated alias. Rules in free text go in the project's CLAUDE.md, AGENTS.md or GEMINI.md, above the Nirvana block that nrv init rewrites; your agent reads them and carries them into the briefs it writes. Asking is enough: "in this project, use detailed briefs", "add to this project's CLAUDE.md: every deliverable in Brazilian Portuguese".

Why "the work is done" means something here

An orchestrator can announce anything in its final message. Nirvana-OS backs each delivery with things you can open on disk.

  • Audit trail. Every dispatch, gate verdict and delivery is an append-only event in ~/.harness-logs/<date>/audit.jsonl. Every --exec run also leaves a record in the project's .nirvana/run-kernel.sqlite. nrv audit-view <project> shows the chain.
  • Checked on disk. verify-deliverable compares what the brief promised with the files that exist, and the quality gate runs rubrics per file type, with correction rounds when it fails. A run whose serious failures survive the corrections is withheld, not delivered.
  • Briefs with criteria. Each brief carries the user's request verbatim and observable "done when" criteria; the ones marked (blocking) decide whether the delivery passes.
  • Spend under your control. No cost ceiling is set unless you ask for one ("cap each run at US$ 5" sets budget.default_max_cost_usd).

Command reference

For agents and power users. nrv help lists everything; nrv <command> --help gives each command's options. Full reference: docs/CLI.md.

Command What it does
nrv doctor Health check: runtimes, hooks, skills links, dependencies, registries, licenses
nrv init <dir> Make a folder a Nirvana project (--orchestrators=always, --scope=project|merge|global)
nrv list-businesses / list-squads / list-clones Browse the three registries
nrv find "<need>" / nrv search "<topic>" Ranked candidates for a request / keyword search across the libraries
nrv brief template|templates|check|decide|init The brief the orchestrator writes for one business, from a brief template
nrv dispatch <business> --brief-file <file> --exec Run a business on a brief; --squad <slug>[:<capability>], --agent-x or --auto name other targets. Refuses without --exec
nrv run-track status|stop <run> Read a run's state, or stop it cleanly
nrv revise <project> "<change>" Change a delivery in the same runtime session
nrv audit-view <project> The audit chain of a run
nrv activate <squad>|--all Prepare a squad's environment in ~/.nirvana/envs/<squad> ahead of time
nrv config list|get|set|unset|explain Settings, with --global or --project
nrv update [--check] / nrv update <pack> Update the engine / a pack
nrv validate <squad|business|clone> <slug> [--fix] Admission gate for one entity
nrv glance Web cockpit (--read-only to browse only)
nrv uninstall --engine Remove the engine and keep your content

FAQ

Do I need to know how to code? No. You describe outcomes in plain language; your agent runs Nirvana, and the businesses and squads write, check and deliver.

Does it replace my agent? No. It runs on top of it. Dispatched work runs on the runtime of your session unless you name another: Claude Code, Codex, Antigravity, Pi, Kimi, Grok, Qwen Code or OpenCode.

Where does my work live? On your machine: content in ~/businesses, ~/squads and ~/businesses/_library/dna, deliverables in each project's outputs/. There is no Nirvana cloud in the loop.

What if nothing in my library fits? The request still runs, on agent-x, the generalist. You can also ask your agent to create the business or squad that is missing.

Windows? Native, through Bun, with no WSL needed. If you prefer WSL, install everything inside Ubuntu, your agent included.

Troubleshooting

  • PowerShell shows >> and waits. A quote was left open in the paste. Press Ctrl+C or open a new window, then paste the single line again.
  • macOS prints "Saving session… [Processo concluído]" after you close Terminal. That is Terminal closing the shell, not an error.
  • The install returns 404. The tag is missing its leading v, or that release does not exist.
  • A pack update fails with a license error. Download the pack again from squads.sh and run its setup.
  • Anything else. nrv doctor checks the whole install and nrv install --check checks the hooks and the PATH. More in INSTALL.md.

License, authorship, and status

Author: Luiz Gustavo Vieira Rodrigues (gutomec / Prospecteezy). No co-authors.

License: the Nirvana-OS Sustainable Use License (SUL) v1.0. The source is published and openly readable, and the engine is free to use. It is source-available, not an OSI-approved open-source license, and certain commercial uses require a separate commercial license. Read LICENSE before relying on any summary, including this one.

Status: beta (0.x, currently 0.15.1). The engine works today and installs in minutes. Expect the surface to keep moving until 1.0.