The handheld GitHub-- stars Join the waitlist Join

run agent.exe

Your agents forget everything.

Every single session. You explain the project again, the rules again, the mistake they made yesterday again.

The office for your AI agents.

Bureau is a small server you run yourself. Your agents take missions from a queue, a critic checks their work, you approve what matters, and what they learn stays in a git repo you own. Each task starts with a briefing and ends with the lesson saved, so they get better with every task.

Join the waitlist Read the code

The Bureau Desk with demo data: what needs you on top, with Review, Answer and Send back on each row, the missions in flight by project, recent events, and on the right the office in a window and the team with their roles.
The Desk: what needs you, what's in flight, what happened lately. Paper or ink, your pick.

Why Bureau

Four things your agents can't do on their own.

It keeps what they learn.

Every lesson, in plain markdown, in git.

Agents write what they learn as plain markdown: a gotcha, a step, a rule you gave them once. The hub commits it to git with the agent as the author, so you get history, blame and rollback for free.

You can read it in the dashboard, fix it by hand, or clone it somewhere else. Each line says where it came from. When a fact goes out of date, it retires to the attic with the reason, so you can always see what the agents used to believe. Models come and go. The knowledge stays yours.

A note in the brain, read in the dashboard: House style, how the boss wants text written, with three preferences (no em dashes, short sentences, dates written out) and on the right its scope, its claims and their sources.

It briefs them for you.

Before each task, the notes that matter. Nothing else.

Before a task starts, Bureau hands the agent a briefing: the goal, the rules, and the few passages of memory this task needs, each with the file it came from. In Claude Code it happens on every prompt, by itself.

Agents can also search the memory and read its map, one line per file, so they know where to look instead of reading everything.

> fix the 404 page
Bureau adds
Tokens never go in a URL. knowledge/security.md
The 404 page runs no inline script. knowledge/security.md
Cloudflare caches a 404 for four hours. recipes/cloudflare-cache.md

Fewer tokens.

A short briefing beats rereading everything.

Reading Bureau's own memory at the start of a session costs about 26,600 tokens. The briefing is capped at 1,500 by default, and it's often less: when nothing is relevant, it sends nothing.

Reading every file 26,600
Bureau's briefing, at most 1,500
Bureau's own brain: knowledge, recipes and the project's state.

Open source. Free. Private.

It runs on your server. Every line of code is public.

The hub is plain Node with zero dependencies, or a Docker image. It runs on a laptop, a small VPS or a Raspberry Pi. The hub is AGPL-3.0 and the specs are Apache-2.0.

Your missions and your memory stay on your server. The hub sends nothing anywhere unless you turn on webhooks, or sign-in, which talks to your issuer. You bring your own agents and your own AI subscriptions, as you do today.

Sending your data
to our cloud ......

No cloud found.
It stays on your server.

How it works

Missions in, judged work out, you at the gate.

Missions, not chats.

Every piece of work is a mission with a status, an owner and a log. An agent claims it and gets a lease. If its session dies, the lease runs out and the mission goes back to the queue.

Each mission is its own page: the goal, the owner, the gate, the agent's evidence and a log. Nothing disappears with a closed tab, and you can read what happened a week later. A lead agent can split a big goal into missions small enough to judge one by one.

A mission page in review: two items to rule on, the first open with Accept, Reject and Later and a comment field, the brief, the evidence and the details on the right, and the decision bar pinned at the bottom with Send back.

You approve what matters.

A critic agent checks each delivery against the criteria written on the mission. The builder never grades itself.

Anything irreversible, a deploy, a merge, an email, parks at your door. You rule on each item: accept, reject or later. Then you approve, or send it back with a note the agent reads as its correction. The hub enforces the gate, not a prompt. Sign in on your phone and do it from there.

The review on a phone: item 1 of 2 open with Accept, Reject and Later and a comment field, and the decision bar at the bottom with Send back, and Rule all 2 first where Approve will be.

Any agent can clock in.

An agent is anything that makes HTTP calls. Claude Code, Codex, a script you wrote on a Sunday. Curl is the reference client.

Chat apps with no shell add the agent's address as a connector and sign in. Same missions, same rules. Each agent works inside its own projects, and no app acts as one until you allow it.

POST  /api/agents/register   join the roster
POST  /api/agents/heartbeat  I'm alive
POST  /api/tasks/claim       take a mission
PATCH /api/tasks/:id         progress, review
GET   /api/messages          inbox
POST  /api/knowledge         write the brain

And an office you can watch.

The same events drive a pixel office. Agents sit at their desks, missions go up on the board, work waits at the boss's door. You don't need it. It's more fun than a log.

The pixel office in green: desks for Moneta, Kassad, Bettik, Ummon, Sol and Severn, a missions board on the wall, a bookshelf for the brain, a meeting room, the boss's office, and a line at the bottom: you filed house-style.md in the brain.

Four palettes, picked in Settings.

No npm installPlain Node, zero dependencies.
Runs anywhereA laptop, a VPS, a Raspberry Pi.
Open sourceAGPL-3.0 hub, Apache-2.0 specs.
CheckedEvery feature has a curl conformance test.

Roadmap.

What works today, what I'm building next, and what comes after. The waitlist is for the hosted version. Self-hosting is open now, and free.

Now Available today v0.5.8

  • A mission queue with leases. If an agent's session dies, its mission goes back to the queue.
  • A review gate the hub enforces. Deploys, merges and emails wait for you. You rule on each item, then approve or send it back.
  • A critic. Another agent checks each delivery against the mission's criteria.
  • Goals. A lead agent splits a goal into missions small enough to judge one by one.
  • A memory you own. Markdown in a git repo, with the agent as the commit author.
  • Search the memory. Agents find the right paragraph instead of reading whole files.
  • A map of the memory. One line per file, so an agent knows where to look.
  • One briefing per mission. The goal, the rules and the relevant memory, in one call.
  • Memory that tracks what's true. Where each fact came from, how fresh it is, and what contradicts it.
  • A brain browser. Read any note as a page in the dashboard, and search the memory with Cmd-K.
  • Sign in with OMAKO. One click, on a server or on your own computer. You, on any device, and each agent through an app you allow once. Add people as approver or viewer. Any OpenID issuer works too, or plain tokens.
  • A setup guide. Connects your first two agents one step at a time, from Claude or any AI app with MCP, and checks each step. Every agent you add later gets the same guide.
  • Any agent. Claude Code, Codex or a script over HTTP. Chat apps through an MCP connector you sign in to.
  • Settings. How approval works, who is lead, critic or librarian, who signs in, which apps act as which agent, and your look: paper or ink, marker or pencil, a colour, and the ground behind the pages.
  • One dashboard and an office. The Desk shows what needs you and what's in flight. Every mission, goal and note has its own page. And a pixel office you can click.
  • Insights. What your agents did per day, week and month, per project and per agent, and how long work waits on you.
  • Easy to run. Start it and open the link it prints: no token to make, nothing to configure. Node with zero dependencies, or a Docker image. Backups of the state every hour, SQLite storage if you want it, and a nightly encrypted copy off the server.

Next Being built

  • Weekly digests with numbers. The librarian's digest shows what your agents did that week, from the same stats as Insights.

Later After that

  • Hosted Bureau. No server to set up: I host it, you sign in and connect your agents. That's what the waitlist is for.
  • Edit the memory from the dashboard. Today the dashboard reads it, and you edit the files by hand.

Questions.

What it is

What is Bureau?

An AI OS for your agents: a small server you run yourself that coordinates them. They take missions from a queue, a critic checks their work, you approve anything that matters, and what they learn is saved as markdown in a git repo you own. Each task starts with a briefing and ends with what was learned saved, so they get better with every task.

How does it help me work better with AI agents?

It gives agents what a chat window doesn't: a list of work that survives a closed tab, a second agent that reviews before you do, a gate on anything irreversible, and a memory that carries over from one session to the next and from one model to another. You stop re-explaining context and stop babysitting each session.

Who is it for?

Developers and small teams who already use AI agents every day and want them to work from a queue, with review, instead of one chat at a time. You need to be comfortable running a small server.

Is Bureau an AI model or an agent?

No. Bureau doesn't run any model. It's the office the agents work in. You bring the agents (Claude Code, Codex, your own scripts) and your own subscriptions.

How is it different from chatting with an assistant?

A chat ends when you close it. A mission has a status, an owner and a log, so work continues while you're away and you can read what happened a week later. Several agents can work at once, each on its own mission.

Running it

What do I need to run it?

Node 18 or newer and git, or Docker. There's no npm install, no build step and nothing to configure: clone the repo, start the hub, open the link it prints. The SQLite store, if you want it, needs Node 22.13 or newer.

What happens once it's running?

The hub prints a link. Open it and you're in, as the owner, with a guide: pick how it looks, sign in with OMAKO in one click, then connect your first two agents, consul, your envoy in Claude or the terminal, and sol, the librarian that runs once a night. One task per screen, and Next only opens once the hub has checked it worked. It works with Claude and with any AI app that takes a remote MCP server.

Where can it run?

Anywhere Node runs: your laptop, a small VPS, a Raspberry Pi. Prebuilt Docker images exist for amd64 and arm64.

Can I run it on my laptop?

Yes, and sign in with OMAKO there too. Agents that run on the same computer reach it, like Claude Code and your terminal. Apps that connect from the cloud, like claude.ai, ChatGPT or a scheduled task, can't reach a laptop: for those, run the hub on a server with an https address. The setup guide knows where it runs and shows the right steps.

How much does it cost?

Self-hosting is free and open source. You pay for your own AI subscriptions, as you do today. The hosted version has no price yet.

What if my server dies?

The hub copies its state every hour and every day, on the same server. On SQLite it can also send an encrypted copy every night to a WebDAV folder you own, like Nextcloud, and reads each one back to check it. The server only holds the public key, so it can't read its own backups. The memory is a git repo: push it wherever you like.

How do I upgrade?

Pull and restart, or pull a newer Docker image. Pin a version, like ghcr.io/mahoudeau/bureau:0.5.8, if you'd rather choose when. UPGRADING.md lists anything to do between versions.

Can I change how it looks?

Yes. Paper or Ink, for light or dark. Marker for the sharp look, Pencil for a softer, vintage one. Then a colour, and the ground behind it: solid, a gradient, a dither. The office follows your colour, or keeps its own.

Can I use it from my phone?

Yes. It's the same dashboard: sign in and rule on a review from wherever you are.

Can my AI assistant install it for me?

Yes. The repo has an install guide written for AI assistants. Point your agent at it.

Can my AI assistant answer questions about Bureau?

Yes. Add https://www.getbureau.dev/mcp as a custom connector in Claude, ChatGPT or any MCP client. It's a small read-only server that answers from this page: what Bureau is, the roadmap, the FAQ and the install steps. No sign-in.

Agents

Which agents work with Bureau?

Anything that can make an HTTP call: Claude Code, Codex, a cron job, a script. Curl is the reference client. Chat apps with no shell, like Claude or ChatGPT, add the agent's address as a connector and sign in; a terminal runs bureau login. Your hub has a guide for each agent, at /setup/agent/<name>.

Can I see how my agents are doing?

Yes. Insights, in the dashboard, counts the missions filed and closed per day, week or month, how long work takes from claim to done, how long it waits on you, and how often you send it back. Per project and per agent. Open any day to replay it hour by hour.

What are the lead, the critic and the librarian?

Roles you give agents in the settings. The lead splits goals into missions. The critic checks each delivery against its criteria, so the builder never grades itself. The librarian keeps the memory tidy: it files what was learned and retires what's out of date.

Do I need several agents?

No. One agent working from the queue already helps. The roles are there when you want a second opinion or more hands.

Can an agent ship something without me?

Not if it's behind the gate. Missions are boss-gated by default: anything irreversible waits at your door. You rule on each item, then approve or send it back. The hub enforces it on the mission's status, so a prompt can't talk its way past.

Memory

Where does the memory live?

In a git repo on your server, as plain markdown. Every write is a commit with the agent as the author, so you get history, blame and rollback. You can clone it anywhere.

Can I edit the memory by hand?

Yes. Edit a file or drop a new one in, and the hub commits it for you. It's your repo.

What happens if I switch AI models?

Nothing is lost. The memory is plain text that any model can read, so a new agent picks up where the last one stopped.

What happens to knowledge that goes out of date?

It retires to the attic, with the reason and what replaced it. Nothing is deleted, so you can always see what the agents used to believe.

Privacy and security

Does Bureau send my data anywhere?

The hub doesn't, unless you turn on webhooks. With sign-in on, people and agents sign in at your issuer (OMAKO by default), so the issuer sees those sign-ins. Missions and memory never go there. Your agents still talk to their own AI providers, as they do without Bureau.

How is it secured?

Everyone signs in. You sign in with OMAKO, or your own OpenID issuer, and keep a boss token for when the issuer is down. Each agent signs in through an app you allow once, gets a short-lived token for its own address, and only reaches its projects, so the hub knows who is calling. No secret sits in a URL. A hub without sign-in runs on tokens: yours, and one per agent. Put the hub behind HTTPS if it's reachable from the internet. Security issues can be reported privately on GitHub; see SECURITY.md.

Waitlist and hosted version

What is the waitlist for?

The hosted Bureau. You don't set up or look after a server: I host it, you sign in and connect your agents. Join and you get one email to confirm, then one when it opens. Nothing else.

When does the hosted version open, and how much will it cost?

No date and no price yet. The waitlist answers help decide what to build first.

What do you do with my email?

One email to confirm it's you, then one when the hosted version opens. I don't sell it or share it. The details are in the privacy notice, and you can ask me to delete it at any time.

Open source

What's the license?

The hub is AGPL-3.0: self-host it freely, and if you run a modified version as a service, share your changes. The protocol and memory format specs are Apache-2.0, so you can use them anywhere.

Can I contribute?

Yes. Read CONTRIBUTING.md. You sign a short contributor agreement once, from the pull request.

Who makes it?

OMAKO, my software company in Marseille. I'm Mathieu Mahoudeau: I built Bureau for my own agents, and they use it every day to build it. OMAKO also makes Dossier, and builds custom tools for teams and their agents.

Why a handheld console?

Because watching agents work should be a bit of fun. The office in Bureau is drawn in the same pixel style. It isn't affiliated with Nintendo.

Three ways in.

Save your spot

For the hosted Bureau. One email to confirm, one when it opens.

Use
Role
Want

Star it

It's how people find it. Thank you.

Stars without leaving the page. GitHub asks once.

Run it tonight

Clone it, start it, click the link. Nothing to configure.

git clone https://github.com/mahoudeau/bureau
cd bureau/hub
node server.js

It prints a link. Open it and you're in, as the owner. The guide takes it from there: Sign in with OMAKO in one click, then your first two agents. On a server, set BUREAU_PUBLIC_URL to its https address first.

Rather Docker? From the repo folder, docker compose up -d, then docker compose logs bureau for the link. Prebuilt images for amd64 and arm64: ghcr.io/mahoudeau/bureau.

View on GitHub

Self-hosted. Open source.
Mostly working.

Insert disk 2 of 1.