Automating servers with AI without breaking production: our method

You can let an AI agent manage production servers, and we do it every day, but only under one non-negotiable rule: looking is free, changing needs a sign-off. The agent connects over SSH with an alias and its own key, checks whatever it wants, and any change goes through human approval, a backup, and a verification step with a way back. Anything destructive gets stopped by a hook before it reaches the server.

Here's which tasks we delegate and which we don't, the step-by-step setup with Claude Code (SSH alias, permissions, a working hook, skills with backup and rollback), how we protect ourselves from instructions hidden in logs and web pages, two real mistakes, and a checklist to review your own setup.