brAIn

What it is

A Home Assistant add-on that runs Claude Code and a suite of tools inside HA, which builds a permanent memory of your house.

It sees the whole system — every entity, device, area, floor, dashboard, helper and automation — and it can change any of it. Explain a broken automation. Fix it. Write a new one. Remember why, next time.

That memory isn’t a black box. Open it, read it, edit it, correct it. An insights panel shows what it knows about your house and what it’s done there — in the sidebar, or embedded straight into your dashboards.

Reach it however you want: as your conversation agent, through a full-featured chat interface, or from native Claude Code. Your automations can call it too — which means your house can ask for help before you notice anything’s wrong.

One install, one sidebar panel, one login. Runs on the Claude Pro or Max subscription — or your own API key.

It runs Home Assistant

Not just “turn on a light” — 65 admin services reach the parts of HA that normally live behind the Settings UI, plus 36 native tools and a real shell in /config. What it can reach →

It finds what's broken

A dying battery, a sensor that stopped reporting three weeks ago, an automation whose trigger can never fire. Press Fix it and it makes the change. Findings →

It remembers your house

One plain-markdown document of durable facts — nicknames, rhythms, the devices that are meant to behave oddly. Everything brAIn does reads it, and brain memory export carries it to new hardware — no starting over. Memory →

It explains your house

Insight cards with real interactive charts, proposed for your home rather than shipped as defaults. Ask anything, get a card. Insights →

It talks, in a voice you choose

Pick brAIn in Settings → Voice assistants. Give it a personality, or run several agents with different names and permissions. Voice →

It has a terminal

The real Claude Code CLI with your /config in front of it, as a chat or as a true terminal — same session, one button between them. Permission asks arrive as cards in the chat, it picks its own model, and conversations delete with an Undo. Terminal →

It works from automations

brain.run_task for background jobs, scheduled insight reports rendered to sensors, and events your automations can trigger on. Automations →

It can be undone

Every file Claude writes to is snapshotted first. brain undo lists the edits in plain English and puts one back. The CLI →

It administers your setup

Most AI integrations can turn on a light. brAIn runs the installation. It reaches Home Assistant three ways at once — native tools for reading and controlling, admin services for the registry, and a shell for everything that is still a YAML file.

• Organisation — create, rename and delete areas, floors and labels; set icons and voice aliases; move devices and entities between them. “Put every upstairs room on a floor called Upstairs and rename the Shellys to what they actually control.”
• Devices and entities — rename either, change an entity_id, hide, unhide, enable, disable. Find references to things that no longer exist and clean them up — dry-run first, so you see the list before anything goes.
• Dashboards — read, create, rewrite, restore a previous version. It can build a dashboard from a sentence.
• Automations, scripts and scenes — it edits the YAML, validates it, reloads the domain, then reads the traces to see whether the thing actually fired and why it didn’t. That last part is what makes “write me an automation” work on the second try instead of never.
• The house’s own record — history and long-term statistics, the logbook, the error log, your other add-ons, weather, camera snapshots it can actually see, and every service any integration exposes.

Nothing is create-only. Everything that can be created can be renamed and deleted, and every attribute a create_ service accepts has a service that changes it later.

It fixes what it finds

A finding is something wrong with your house, filed by brAIn on its own while it was looking at something else. Each one carries the evidence — the entity, the number, the day it started — and the specific change that would resolve it.

Press Fix it and Claude makes the change, verifies it, and reports what it touched. Press Discuss to talk it over first, changing nothing. I fixed it and ✕ Wrong are endings: the answer becomes a fact in memory, and the row is gone for good rather than joining a pile of dismissed cards. Both take an optional one-line note, and the note is the half that teaches — “that sensor always reads on, it’s not stuck” becomes a standing fact about your house, so the next run doesn’t make the same mistake in new words.

Findings can also reach you without the panel open: sensor.brain_open_findings carries the open count with the texts as attributes, a brain_finding event fires for each new one, and setting one option (findings_notify_service) turns them into phone pushes — no automation needed.

It learns, and everything reads what it learned

Memory is one markdown document you can open and edit. brAIn writes to it from conversations, insight runs and background study sessions — and voice, insights and the terminal all read it. Tell the voice assistant something and the cards know it.

When brAIn believes something it can’t confirm, it says so as a guess you answer with one tap — “The garage fridge is meant to run 24/7 — right?” Three open at a time, never more. Guesses sit at the top of the Findings tab, because “is this broken?” and “have I got this right?” are the same job — a question only you can answer — and one list should hold everything waiting on you. Yes files it as a fact; No records a dead end that is never revisited.

It works from the sofa

The panel is one ingress page with five tabs, built for a phone as seriously as a desktop: every target is at least 44px, no width gets a row of bare glyphs, and the terminal folds the chrome away when the keyboard comes up.

Built for Home Assistant

What makes this different from running Claude Code on a regular machine:

• Your Claude subscription covers it — OAuth sign-in, so Pro/Max includes your usage; an API key works too. One login covers the panel, the terminal, voice, insights and memory.
• 36 purpose-built MCP tools — live states, device control, cameras, history and statistics, registries, dashboards, traces, templates, logs, reloads.
• Auto-generated context — every boot regenerates /config/CLAUDE.md, a snapshot of your install that Claude reads at the start of each session.
• Usage-limit sensors — the same session and weekly utilisation as claude.ai, as HA sensors with reset times, plus a budget that pauses scheduled work before it eats your week. Manual presses are never blocked.
• Safe by default — voice agents get HA tools but no shell, file or web access; per-agent blocked-service lists; a non-root sandbox behind an AppArmor profile (6/6 on the add-on store’s security rating); the terminal port ships unpublished and password-protected; backup_exclude keeps the Claude credential out of HA backups; secrets.yaml is never snapshotted and never edited.
• Two CLI commands, not fourteen — brain for its own faculties, ha for Home Assistant chores. The CLI.

What it doesn’t do

brAIn can’t add an integration (config flows are interactive, so it walks you to the right screen and you click — removing one it can do), never touches credentials, and doesn’t back up your configuration — Home Assistant’s own backups already do that. What it keeps is an edit journal of the files it touched, so brain undo can put one back. The full scorecard: what brAIn can and can’t reach.

Back up Home Assistant first — somewhere that isn’t this machine

brAIn edits your real configuration: automations, dashboards, helpers, entities. It is built to be careful, it snapshots files before it changes them, and brain undo puts them back — but none of that is a backup you can restore from if something goes wrong. Settings → System → Backups, then copy it off the device: a backup that only exists on the machine you are changing is the one you cannot use when that machine is the problem.

Quick Start → — add the repo, install (a prebuilt-image pull, not a local build — a download, not minutes of compiling), restart HA once, sign in, press Start learning. About five minutes, plus a few minutes of study in the background.