The safety checkpoint every AI action passes through — before it happens.
Your agent can send the email, issue the refund, update the record. SafeAct makes sure it does that safely, once, and on the record — so your risk team says yes instead of no.
pip install safeact
$ npm i safeact
The agent isn't too dumb to deploy. It's too untrusted to let act.
Almost every company launched an AI agent this year. Almost none put one into production — not for lack of intelligence, but because no one could trust it to touch real money and real customers.
Four things, every single time your agent acts.
Enforced in code — underneath any agent framework. The model proposes; SafeAct decides whether the action is allowed to happen.
Checked
Every action is checked against your rules — in code, not a prompt the model can talk its way around.
Held when unsure
Low-confidence actions are routed to a human to approve or reject — not fired off blindly. The confidence signal comes from real checks (our companion library wobbly), not the model grading its own homework.
Exactly once
Crash, restart, retry — the refund still goes out once. And if it's genuinely unknowable, SafeAct says so instead of guessing.
Provable
A tamper-evident receipt for every action: what happened, why, which rule was in force, and who approved it.
One refund agent. Five moments. The whole thesis.
Nothing staged — this is the product doing its job, start to finish.
Agent issues a routine $50 refund.
✓ CLEAREDAgent tries a $50,000 refund.
✋ HELDA human clicks approve. It runs.
✓ RANServer dies mid-payment, restarts.
① NO DOUBLEOne button: "Prove it."
🔒 RECEIPTFree core. Paid control.
Two free libraries get developers in the door — the action spine and the verifier. The control plane is what their risk and compliance leads pay to keep.
- Exactly-once, crash-safe execution
- Honest "in-doubt" state — never a blind retry
- Saga rollback for multi-step actions
- Python & TypeScript, ten lines to adopt
- Runs local — one file, no server
- Ships with wobbly — label-free checks that flag unstable AI outputs before you act on them
- Human-approval inbox for held actions
- Rules-as-code policy engine, versioned
- Tamper-evident audit of every action
- One-click "prove it" regulator export
- Roles, multi-team, SSO
Let your AI act. Keep the proof.
SafeAct comes out of running real, action-taking agents in production — where we learned exactly how they break. That knowledge is the product.
Get started free