AgentCo

A virtual company of AI employees that runs on your own machine. You give orders in plain language; a director plans the work, specialists run it in parallel, and everything they produce lands in a folder you own. agentco is a working codename.

You have already felt this

You hand an AI agent something real. It runs for twenty minutes.

And for those twenty minutes you have no idea where it is, what it is doing, how much it has cost you, or whether it has been heading the wrong way since minute three. You find out at the end, when it hands you the wrong thing — and the only record of how it got there is a wall of transcript nobody wants to read.

Then you try to make it useful on your work, and hit the second wall: before an agent can touch your CRM, your invoicing system, your internal API, somebody has to write an MCP server first. So the tool is for developers, and you are not one, or you are one and you have better things to do.

Why this is worth the next two minutes

• It runs on your Claude Code subscription. Not an API key, not per-token billing, and no server of ours anywhere in the path. The daemon on your machine talks to Anthropic directly.
• Your company is a folder. company.yaml, markdown, JSON. No database — the entire storage layer is four runtime dependencies. You can open any of it in Notepad, diff it in git, and back it up by copying it.
• The numbers are measured, not claimed. bench/ reproduces every figure in docs/FINDINGS-sdk-2026-08-14.md, and every PR has to answer a token-cost checklist before it merges.
• 1,200+ tests, and the ones that matter are named after the failure that paid for them.
• Source-available, and each release becomes Apache 2.0 on its own two years later — with the word irrevocably in the licence, so it does not depend on me still being here.

What it actually is

A director that plans and delegates but does no hands-on work, and workers that arrive, do one thing, write a file, and vanish:

   You ──► DIRECTOR (long-lived, plans, delegates, never does the work itself)
                │ TaskBrief (a DAG)
      ┌─────────┼─────────┬─────────┐
      ▼         ▼         ▼         ▼
   Worker    Worker    Worker    Worker      stateless, parallel
      │ Receipt (≤800 tokens, with a human-readable `say` field)
      └─────────┴─────────┴─────────┘
                │
        KNOWLEDGE GRAPH (markdown + frontmatter — readable, greppable, yours)

Workers never talk to each other. Every exchange goes through the director or through an artifact, and that is an economic decision rather than an aesthetic one: agent-to-agent chatter is the largest and least controllable source of token burn in every multi-agent system.

You watch it happen as a plan in four understandable steps. The raw transcript and the per-token cost table are one click away, for the day you want them.

And it can use your own systems without anyone writing a server. Paste an OpenAPI link, paste a cURL command, or fill in a form — then press Test. → docs/SPEC-connectors.md

Run it

Requirements: Node.js ≥ 22 and a Claude subscription. You do not have to install Claude Code separately — the package brings its own copy; agentco login signs it in, once. No API key.

From the clone you are already reading

Nothing below is in the repository — dist/, web/dist/ and node_modules/ are all generated — so a clone has nothing to run until you build it:

npm install                       # server dependencies
npm --prefix web install          # the interface is its own package; there are no workspaces
npm run build:all                 # tsc + vite — this is what creates dist/cli/index.js
node dist/cli/index.js init       # creates ./company here
node dist/cli/index.js login      # signs in the Claude Code this package brought
node dist/cli/index.js doctor     # every line ✓ except the daemon — you have not started it yet
node dist/cli/index.js start

Install the published package

For running it rather than working on it:

npm i -g @agent-co-app/cli

Then, in the folder where you want the company to live:

agentco init
agentco login     # once per machine; skip it if `claude` is already signed in here
agentco start

init creates ./company right there and start opens the interface. npm i -g installs into npm's global prefix, so the folder you are standing in stays empty until init — that is init's whole job.

On macOS or Linux, if npm i -g answers EACCES: npm's global folder belongs to root. Do not reach for sudo — that installs agentco as root while the daemon runs as you, and the Update button inside the app can then never write there, failing after the daemon has already stopped, where no message can reach your screen. Point npm at a folder you own instead:

npm config set prefix ~/.npm-global
echo 'export PATH=$HOME/.npm-global/bin:$PATH' >> ~/.bashrc   # ~/.zshrc on zsh, the macOS default
source ~/.bashrc

agentco update              # newest version, then starts the company again
agentco update --to 0.1.3   # or back to a named one, if a release goes wrong
agentco version
agentco doctor              # checks Node, the company folder, Claude Code, sign-in, the daemon
npm uninstall -g @agent-co-app/cli

Removing the package leaves the company folder alone. It is your data, not the program — delete it yourself if you mean to.

Windows has a second door: a packaged installer at agent-co.app that brings its own Node and adds a Start-menu icon. It updates by installing over itself rather than through agentco update, and the app tells you which door you came in through at the bottom of the Settings panel.

Run it in Docker

Three steps, and only the second involves typing anything unusual.

1. Mint a token — on any machine, including your own laptop. It does not have to be the server, and you do not need Claude Code installed there:

node dist/cli/index.js login --token   # from a clone — after npm install && npm run build:all
claude setup-token                     # or this, if you already have the CLI

It prints the token once and stores nothing. Copy it.

2. Fill in .env:

cp .env.example .env        # copy .env.example .env   on Windows

CLAUDE_CODE_OAUTH_TOKEN=sk-ant-oat01-...
AGENTCO_PORT=7317

3. Launch:

docker compose up -d

Open http://127.0.0.1:7317. That is the whole thing — one container, your data in a named volume, and the port published to loopback only, including when the machine is a VPS.

Details, the tunnel route, and what to know before exposing it to a network: docker/README.md.

Design principles

1. Own the artifact, not the prompt. The value lives in the files inside your company's folder. They survive independently of anything Claude Code changes.
2. An agent is a stateless function. It arrives, does the work, writes a file, and disappears. Memory lives in the knowledge graph, not in the context window.
3. Every token must justify its own existence. Cost is the primary design constraint, not a feature bullet.

What leaves your machine, precisely

Your files, your company folder and your knowledge graph stay where they are; nothing is uploaded and there is no account with us to have one. What does go out is what any Claude Code session sends — the prompts and the file contents an agent actually reads — straight from your machine to Anthropic, on your own subscription. Said plainly because "no data ever leaves your machine" would be a nicer sentence and a false one.

Documentation

| File | Contents | |---|---| | docs/SPEC-token-economy.md | Read this first if you only read one. Token-cost rules, prefix-cache architecture | | docs/SPEC-connectors.md | The signature feature — employees that can CRUD your own REST API / MCP server | | docs/SPEC-2026-08-14-agentco.md | System spec — org structure, roles, the Task/Receipt protocol, the knowledge graph, the scheduler | | docs/SPEC-cli.md | Process model, command set, configuration | | docs/SPEC-deploy.md | Personal machine · Docker · VPS + domain, and what changes between them | | docs/SPEC-packaging.md | The installer, the signed update channel, the release sequence | | docs/SPEC-canvas.md | Node-based canvas — a drag-and-drop office (the next UI generation) | | docs/SPEC-ui.md | The shipped UI, SSE events | | docs/FINDINGS-sdk-2026-08-14.md | Real measurements against the Claude Agent SDK — numbers, traps, and the decisions they led to | | docs/TEST-WALKTHROUGH.md | Walk the product by hand, including the Docker door | | docs/ROADMAP.md | Development milestones | | bench/ | Measurement scripts — every number in FINDINGS is reproducible |

License

FSL-1.1-ALv2 — source-available. This is not open source under the OSI definition, and I don't call it that.

You may do, today: read all of the code · run it for any purpose, including commercial use and internal company use · modify it · fork it · publish your modifications · build an internal product on top of it.

Exactly one thing is forbidden: selling a product or service that competes with agentco. And that restriction has an expiry date — see below.

Every release turns into Apache 2.0 on its own, exactly 2 years later

"We hereby irrevocably grant you an additional license … under the Apache License, Version 2.0 … effective on the second anniversary of the date we make the Software available."

• It cannot be revoked. The word irrevocably is in the license itself. I have no power to take it back — not if I change my mind, not if the project is acquired, not if I disappear.
• The clock runs per release. Publishing a new release does not push the older one's clock back by a single day.
• Result: a sliding two-year window — anything older than that is already fully Apache 2.0, with no button for anyone to press.

📅 Every GitHub release states the exact date it becomes Apache 2.0 — spelled out, not left vague. Vagueness doesn't stop someone determined to copy the code (they can work it out from the license themselves); it only costs the trust of someone on the fence about the project.

Contributing

PRs are welcome — read CONTRIBUTING.md first.

Your first PR needs the CLA signed with one line pasted into the PR description. You keep the copyright on what you write; the license you grant is exactly what makes the "turns into Apache 2.0 on its own" promise above enforceable.

Every PR must answer the checklist at the end of docs/SPEC-token-economy.md — a change that makes any golden scenario's cost worse by more than 10% will not be merged.