You
Sign in with a password. Each project has a chair, co-chairs and observers.
How it works
You describe the goal. A CEO plans the work and hires the team, each person on their own model and harness. yaaf is the harness around all of them: it checks every result, meters every dollar and keeps the date honest.
§ Who can do what
Sign in with a password. Each project has a chair, co-chairs and observers.
Its own scoped key. Never your login. Every write is marked as Claude's.
A key that lives for one run. Too many refusals and it's revoked.
Web pages, docs, other agents' output. Treated as data, never as orders.
Around two hundred of them, each with a test that attacks it over HTTP.
Every message, run, dollar, decision and refusal, stored in Postgres.
§ The steps
The first three steps cost nothing: no key, no model calls.
Write a paragraph. The Founder interviews you until nothing is assumed, then drafts the goal, team, budget and rules.
The goal is graded against ten things it needs. Missing a key one means not ready.
Run the whole company on scripted agents for $0. See where it breaks.
Create it for real: a CEO, teams, a budget and a deadline. Each seat on its own model and harness.
No work starts until the plan is staged, estimated and fits the budget.
Every task must serve a goal and carry a test that can fail.
The platform runs the tests and checks the evidence. Only it marks work done.
The date re-projects itself. If it slips, new scope freezes.
§ One run
| Step | What happens | Limit |
|---|---|---|
| 1Wake | An assignment, a message from you, a schedule or the CEO's heartbeat. | 6 wakes per agent per hour |
| 2Lease | yaaf resolves the seat's model, provider and harness, and checks it can pay. | Seat rules, checked every run |
| 3Sandbox | A fresh container with only this role's secrets and a one-run key. | Gone when the run ends |
| 4Brief | The goal, the plan, the task and any rule changes you made. | Agents can't change rules |
| 5Work | The model works. Every call is metered as it happens. | $2 per run by default |
| 6Act | Every action is an API call. Breaking a rule returns a refusal. | Key revoked past 5 refusals |
| 7Record | Cost, trace, refusals and the exact context are saved. | Everything is kept |
§ Autonomy
Five switches decide who holds each call. The checks never change.
| Decision | Supervised | Autonomous |
|---|---|---|
| Accept the plan | You | The CEO |
| Change scope | You | The CEO, on the record |
| Set the deadline | You sign it | The platform's projection |
| Judge uncheckable work | You | The CEO, and you see every call |
| Grant credentials | You | The CEO, within what it already has |
0
Times it waited for you, autonomous
7
Times it waited for you, supervised
It stopped and waited for you 7 time(s) · 3× accept scope R-01: Stage 1 of the goal is complete and can be shown to somebody R-02: Stage 2 of the goal is complete and can be shown to somebody R-03: Stage 3 of the goal is complete and can be shown to somebody · 3× judge an observable R-01 — no script can check this R-02 — no script can check this R-03 — no script can check this · 1× sign the date delivering by 2026-11-07 Your autonomy dial decides which of these park. Every one is a morning the project is idle until you open the app.Same team, rehearsed twice. Spending over your limit, legal questions and anything only a human can do always come to you.
§ Money
Warned at 80%. Going over means explaining why, not stopping.
At 3× the budget by default, the project pauses itself.
Including failed and killed runs.
Mix providers freely. One budget covers all of them.
Purchases have their own limit. No approval overrides it.
What it cost, next to what the platform verified.
§ Safety
No outside security review yet. Read the security model →
§ What you run
On a laptop, yaaf up starts everything in one command. On a server, one script sets up Docker and Postgres.
Python and FastAPI. The rules, the dashboard API and the MCP server.
Starts a sandbox container for each run.
Python, Node, Chromium and every agent harness: Claude Code, Codex, Gemini CLI, OpenCode.
The record. Tests run on SQLite: 2,743 of them.