Homelab automations
My home automations as small Python scripts: an AI can write them, but I approve each one, and they run in a sandbox that can only do what they declared.
Think Node-RED, but each automation is a short Python script. This service, called astromech, is what my other apps call when something happens, and what calls them back. A script can start from a webhook, a schedule, a Home Assistant event or an incoming email.
What I care most about is control. Each script declares what it's allowed to do (which services it may call, and how), and that's checked from the code before it ever runs. Scripts can come from me or from an AI assistant through the MCP server (MCP is the protocol AI assistants use to call outside tools), but nothing runs until I approve it in the web app. A dry run shows what a script would have done, without doing it.
Scripts run in a sandbox, a fresh process each time, with no passwords or keys of their own. They can only use a small set of functions, and astromech makes the real calls for them.
It's still new: I started it in October, and I'm writing the first real scripts now. It's the same question I deal with at work, at home size: which actions an automated agent can take on its own, and which ones stay with a person.