snoopd sits between an AI agent and my dozen-odd Linux and macOS boxes. It exposes about a dozen fixed, read-only tools: system health, disk, memory, Docker containers, listening ports, cron jobs, kernel errors. Each tool opens an SSH connection from my own machine with my own key, runs one boring, pre-defined command, parses the output and disconnects.
The agent never touches SSH, never sees a key, and can’t pass its own commands through. A health sweep runs across every box every three hours.
Limitation, by design: it can only ask the questions I’ve coded into it. That’s the point.