Ratel

Ratel is an MCP server for securely controlling Docker, consisting of a gateway that AI clients connect to and a small node deployed on each Docker host.

Ratel's permission layering is designed as follows: read operations (list, inspect, logs, stats) and safe write operations (start, stop, restart, pulling images) are granted by default; destructive operations (deleting containers, executing arbitrary commands inside containers, starting new containers) require explicit opt-in each time.

In the demo, the AI diagnosed that the application was in a crash loop due to a misformatted configuration value by listing containers, checking the restart count, and pulling and reading 50 lines of logs, then restarted the container on the user's instruction and checked the new logs to confirm a clean start.

Security design should follow the reflex of "scope it first, build trust, then expand gradually," for example pointing a Docker control tool only at a sacrificial spare host rather than critical cluster assets.