Keel Docs
Issue AI permits before consequential actions. Produce evidence anyone can verify after.
An AI Permit is a pre-execution decision record. It may allow, deny, throttle or require review. An AI Permit is not the same as approval. Only an allow decision authorizes execution.
Keel provides agent control infrastructure. It evaluates policy, budget, routing, delegation and approval before dispatching supported actions, records the result in an AI Permit, and produces route-appropriate evidence for later review.
Create a project, connect supported providers, and define authorization rules in the dashboard. Then connect your application using a supported SDK or API. Keel evaluates the requests your application sends and records their decisions and evidence. Opening the dashboard alone does not control external actions. Start with Quickstart.
Use the docs map to choose the right starting point across product, architecture, and security.
Understand what the public product covers, what it excludes, and which claims are route-specific.
Create a project, add a provider key, and send one supported request through Keel.
Review the public runtime model behind AI Permit decisions, provider dispatch, and evidence.
Read the current security posture, isolation model, request-freshness controls, and prompt-firewall limits.
Recommended path
- Read Overview to choose the right path through the docs.
- Read Product Boundary to understand the active public surfaces and non-claims.
- Run Quickstart to get a governed request through the stack.
- Use Architecture and API Reference to understand the runtime surface.
- Review Security, Threat Model, and Scope and Limits before production rollout.
API reference
Start with API Reference, then use Execute, Executions, AI Permits, Proxy Execution, Idempotency, and Errors for route contracts and examples.