What is yaaf?
Open-source software for building a company of AI agents. You describe the goal and hire a CEO; the CEO runs a team for weeks, and yaaf's server enforces the rules, not the prompt.
FAQ
Short answers. Where the answer is “not yet”, we say so.
Open-source software for building a company of AI agents. You describe the goal and hire a CEO; the CEO runs a team for weeks, and yaaf's server enforces the rules, not the prompt.
Claude Code, Codex, Gemini CLI and OpenCode are agent harnesses: each drives one model with a loop and tools. yaaf puts each of them in a seat on an org chart and is the harness around the whole team: who is hired, what's next, what it costs, when it's due and whether it's really done. Harnesses and models →
No, and there isn't going to be one to upsell you. yaaf is Apache-2.0 software you run yourself. Your agents pay your model provider directly. No account, no telemetry.
Anyone handing off work that runs unattended for weeks and must be checked. Not for quick tasks; use a skill for those.
Every task has a test. The platform runs it, or fetches and checks the evidence. Agents can't mark anything done.
Your choice. Five switches decide what you approve. An example team needed you 0 times, or 7 with approvals on. See both →
Every run is metered live. Runs have a cap, loops are braked, and the project pauses at a hard limit.
No. The date comes from measured pace. If it slips, new work freezes until scope is cut or the date is reset.
Claude, GPT, Gemini, Bedrock, Vertex, Azure, DeepSeek, Mistral, Kimi, Qwen, Grok, Groq, OpenRouter, a local Ollama or LM Studio, or any OpenAI-compatible server: 23 providers through 6 harnesses. Roles name a class (frontier, balanced, fast), so a team runs on whichever provider you pay. Harnesses and models →
For personal tinkering only. Anthropic's terms don't allow running fleets on a personal plan. Use an API key.
yaaf itself costs nothing. The models cost what your provider charges. Nobody has published a real month's bill yet; the $0 rehearsal tells you the most a run can cost and still fit your budget, and every run is metered against a cap you set.
Yes: Ollama, LM Studio or any OpenAI-compatible server. The prompts were tuned on Claude, so quality may vary.
Yes. Clone the repo, install it into a virtualenv (python3 -m venv .venv && .venv/bin/pip install -e .) and run yaaf rehearse examples/01-your-first-hour.json. A whole project runs on scripted agents: no key, no Docker, $0.
On your own machine, a checkout and yaaf up, which starts everything and prints the address. On a server, a Linux box with Docker; one script does the rest. Install guide →
Yes. yaaf is an MCP server. Claude Code, Claude Desktop and Cursor connect with a token. claude.ai needs a public address.
Agents run in sandboxes with limited secrets, but they can reach the internet. No outside review yet. Security model →
Pre-1.0. First commit August 2026. 2,743 tests. One maintainer.
Yes, openly, with Claude Code. About 97% of commits say so.
Try to break a rule. Every rule has a test, and a failing one is the best contribution. Contributing →
Question not here? Ask on GitHub