hook-manager

Filesystem path ownership enforcement restricts autonomous AI agents to writing only within explicitly designated directories, preventing unauthorized modifications and maintaining project security boundaries. It uses generated per-task policies to define these access controls.

MCP server allowlisting is a network access control mechanism that dictates which external services AI agents are permitted to communicate with, ensuring agents only call explicitly authorized servers and preventing unauthorized network requests.

Generating per-task security policies requires accessing the agent registry, hook configurations, and worker manifests to determine appropriate filesystem boundaries and MCP server access controls before an agent begins executing its assigned tasks.

Yes, you can restrict an AI agent's write access to specific directories by applying path ownership enforcement, verifying that the agent only modifies designated folders such as a backend directory and preventing writes elsewhere.

Yes, an agent registry is required alongside hook configurations and worker manifests, as these components provide the necessary framework context to accurately determine and apply appropriate per-task security boundaries.