Documentation

Secure agent actions with one control plane.

AgentSecurity sits between supported agent tools and execution. Your code defines agent identity, tools and intended permissions; the service authorizes exact operations and sends agents, policies, decisions and incidents to one dashboard.

What it protects: the action boundary. If an agent is manipulated, tools still require the policy assigned to that agent.

Choose an integration

LangChain
Use middleware with create_agent.

LangGraph
Install middleware before graph compilation.

CrewAI
Wrap local BaseTool execution.

OpenAI Agents
Attach lifecycle enforcement to local FunctionTools.

Core workflow

  1. Generate a scoped key in the dashboard.
  2. Create one reusable AgentSecurity object.
  3. Wrap the agent where it is built.
  4. Run the application; discovery and policy sync are idempotent.
  5. Review runtime decisions and approvals in the dashboard.

Security boundary

Local interception cannot protect an unwrapped copy of a tool or credentials left directly inside the agent process. High-risk operations should use gateway enforcement so business credentials remain server-side.