yaaf

My open-source platform for handing real work to a team of AI agents, with rules enforced by the server so the agents can't talk their way around them.

Hand real work to a team of AI agents. Keep the final say.

Describe a project in a paragraph. yaaf hires a CEO and specialist teams, gives them a budget and a deadline, and keeps them working for days or weeks. You’re the chairman: you get a daily letter and only the decisions that need you.

FreeApache-2.0Self-hostedNo telemetryMCP serverAny model
localhost:24247/app · Read The Filings
The yaaf project cockpit: two decisions waiting on the chairman, each with the CEO's recommendation and what happens if nobody answers, above a timeline showing two of three outcomes verified and $10.03 of a $60 monthly budget spent
The cockpit. Two decisions wait on you, each with the CEO’s recommendation and what happens if you say nothing. Below: progress against the promised date, and money against the budget.
Why I built it

I spent a year writing about why agents break. Then I built the fix.

Every failure below is in my guides. In yaaf, each one is a rule the server enforces, not a line in a prompt that a clever model can talk its way around.

The failureAgents mark their own work done.
The rule

“Done” means verified. The platform runs the tests itself, and an agent’s word is not enough.

The failureThey invent busywork and call it progress.
The rule

Work has to point at an outcome. Without outside proof, new internal-only work is refused until evidence lands.

The failureThey loop with each other at full price.
The rule

Every run is metered. At the budget limit, the project pauses itself and tells you.

The failureThey can’t tell you when they’ll be finished.
The rule

The date comes from measured pace. If it slips, new work freezes until you cut scope or reset the date.

An agent that breaks a rule doesn’t get a stern reminder. It gets HTTP 403

How a project runs

From a paragraph to weeks of supervised work

  1. Describe it

    Write what you want in your own words. From the dashboard, the CLI, or Claude.

  2. yaaf drafts the team

    A CEO, specialist teams, a budget and a delivery date. You approve or change any of it.

  3. Rehearse for $0

    Scripted agents run the whole plan, so you see where it stalls and what it costs before spending a cent.

  4. Run it as chairman

    Agents work for days or weeks. Each morning you get a letter: what happened, and what needs you.

A look inside

One agent to talk to. A full record of everything else.

The Team page: the chairman at the top, the CEO beneath, and two teams with what each agent is doing, which model it runs on and what it has cost this month
Team. Who’s working on what, which model each seat runs, and what each one has cost.
Chat with the CEO: a question about which fiscal year to use, with options, a recommendation and a deadline after which the default applies
Chat with the CEO. Questions come with options, a recommendation, and what happens if you stay silent.
The task board with Ready, In progress, Being checked, Stuck and Done columns; done tasks carry an outside proof badge
Work and proof. Nothing reaches Done without passing a check. The board is read-only; to change the plan, you tell the CEO.
What you get

Built for work that runs unattended

A CEO that plans the work

It picks the next task, assigns it, and checks the result while you’re away.

Any model, per seat

Anthropic, OpenAI, Google, Mistral, DeepSeek, a local Ollama and more.

Your choice of agent runtime

Claude Code, Codex, Gemini CLI, OpenCode, or yaaf’s own loop.

Works from Claude

yaaf is an MCP server: steer projects from Claude Code, Claude Desktop or claude.ai.

A full record

Every message, run, dollar and decision is kept.

Yours to run

On your laptop or your server. No account, no hosted service, no telemetry.

Quick start

Pick one

Then open localhost:24247/app, add a model key under Settings → Providers, and create a project. Full steps are in INSTALL.md.

Start here · $0

Rehearse it

No server, no key. Python 3.12+.

$ git clone https://github.com/dominicplouffe/yaaf
$ cd yaaf
$ python3 -m venv .venv
$ .venv/bin/pip install -e .
$ .venv/bin/yaaf rehearse \
    examples/01-your-first-hour.json
Recommended

On a server, with Docker

The way to run it for real.

$ git clone https://github.com/dominicplouffe/yaaf
$ cd yaaf
$ deploy/deploy_single_server.sh \
    --admin-email you@example.com
Optional

Connect Claude

Start with the read scope; add steer, create or configure later.

$ yaaf login
$ yaaf mcp-token mint my-laptop \
    --scopes read
$ claude mcp add --transport http \
    yaaf http://localhost:24247/mcp \
    --header "Authorization: Bearer ymt_…"
How it works

The rules live in the server, not in the prompt

The server holds every rule and decides when an agent should wake. The runner starts each agent run in a fresh sandbox. Agents act only through the server’s API, with short-lived tokens, so a rule can’t be argued with. It can only be met.

Good to know before you start

It’s young: the first commit was in August 2026, built in the open with Claude Code.

  • It’s pre-1.0. APIs still change.
  • It costs real money. Agents pay your model provider directly, with your key. Rehearse first.
  • It’s not for quick tasks. For a one-off job you watch, a single agent is cheaper. yaaf is for work that runs unattended for days or weeks.
  • Mind the machine. Read SECURITY.md before running agents somewhere that matters.
The thinking behind it

Read the guides yaaf came from

Try it

Rehearse a project tonight. It’s free.

If you try it, I’d like to hear what breaks. Open an issue on GitHub, or tell me directly.