An AI assistant is useful during an incident when it can inspect current status, recent failures, latency and certificate warnings. It becomes dangerous when the same connection can delete monitors, rotate credentials or act on untrusted response bodies.
Warden exposes a role-scoped Model Context Protocol endpoint. A viewer can investigate and run an immediate check. An editor can also create and organize monitors or pause and resume them. MCP deliberately excludes monitor deletion, user administration, credentials, notification configuration and autonomous remediation.
Use a viewer key by default
Failed HTTP responses may contain text controlled by the monitored service. That makes the response body valuable for diagnosis and also an untrusted input. With a viewer key, a prompt hidden in that body cannot call write tools because the server does not expose them to that role.
Use a separate editor key only for supervised configuration work, then return to viewer access. Give each assistant or machine its own key so it can be revoked independently.
Keep the boundary on the server
Do not rely on a prompt that says “never delete production.” Enforce capability limits in the monitoring API. Serve MCP over HTTPS, keep bearer keys out of repositories and disable the endpoint entirely with MCP_ENABLED=false when it is not needed.
The Warden MCP setup guide lists every exposed operation and the data that may leave the server. That explicit contract is the difference between an AI operating interface and unchecked automation.