Agent integration audit

Use for an authorized investigation of an agent integration. Determine the actual runtime, installed source/version, active configuration and consumer interface; a manifest or another agent's compatibility file does not prove registration. Distinguish declared, registered, reachable, previously observed and currently delivered. Trace each event through prerequisites, debounce/cancellation, state and final effect; a loaded plugin or test fixture does not prove a live notification or external delivery.

For MCP, inventory and validate every actual server entry, authentication source, executable, transport and applicable allowlist. Avoid printing credentials. Pattern scans cannot prove absence of secrets, direct argv is not inherently shell execution, and unpinned packages include omitted versions as well as mutable tags. Treat unsupported schemas, malformed entries or missing approval policy as unresolved; never run untrusted servers as a read-only check.

For hooks, check source and side effects before execution; choose synthetic fixtures and native runtime introspection as authorized. Separate static syntax, isolated behavior, registration and actual destination read-back. Preserve native approval and trust decisions; never automatically answer human prompts or reroute a denied action. Obtain agent-specific details from its supported public interface and current official documentation rather than assuming schemas or paths are interchangeable.

For a read-only evidence map and runtime-specific lifecycle investigation, see plugin and hook evidence. Distinguish observed history from current liveness and never inspect credential values.