How do developers integrate AI Control?

Developers call POST /api/v1/authorize (or aiControl.authorize() in an SDK) before any consequential action, passing the agent, capability, resource and context. The response contains a decision, reasons, risk score, matched policies and an audit ID. Real-time events stream over WebSocket and signed webhooks.

Frequently asked questions

What should my agent do if the decision is not ALLOW?

Never execute. Hold for approval on REQUIRE_APPROVAL, apply the returned limits on LIMIT, and stop on BLOCK or QUARANTINE.

Is there a sandbox?

The backend runs locally with an in-memory driver and demo data — no database required.