Build and maintain a whole Home Assistant setup from inside the box

Marcelo installed Claude Code in the Advanced SSH and Web Terminal add-on on a Home Assistant Green, with init commands that reinstall it on every boot and keep its login, memory and skills in a folder under /config. The /config folder is a git repo. He talks to it in Portuguese, mostly from the terminal in the Home Assistant app on his phone. It edits YAML, runs ha core check, reloads, reads logs, entity state and automation traces, and talks to Home Assistant over the websocket and the official MCP Server integration. It also renames devices in the registry, patched a community card, wrote 17 custom animated cards in plain JavaScript, and generates the whole dashboard from a Python script. A CLAUDE.md holds house-wide rules and it keeps 25 topic notes it wrote itself, read before touching each subject.

Since June 2026 the repo has 580 commits, 129 automations, 73 scripts, 147 devices and 1,163 entities, and a dashboard of 10 tabs and 60 pop-ups built by about 7,600 lines of Python, also output in Spanish and English. He says he barely typed any YAML. It set up Matter so a CO2 sensor no longer depended on the Samsung cloud, ran a security sweep that closed MQTT ports and removed a public page, and traced a phone certificate error to the ISP intercepting Cloudflare traffic during football matches. Mistakes cost real things: one script guard bug disarmed the alarm five times with the house empty, and UI edits were wiped by the generator five times before a drift check was added.

Add-on containers are rebuilt on restart, so anything installed there disappears unless it is reinstalled at boot, and tools are kept to the Python standard library for the same reason. Make /config a git repo before the first prompt, keep secrets.yaml off-limits, and ask for proof from logs and state, not a claim that it should work.