Talking to your house: Claude, ChatGPT and the ha-mcp server

What MCP is, in one paragraph

MCP is an open standard that lets an AI application talk to outside systems through a defined set of tools. The AI client, such as Claude Desktop or ChatGPT, connects to an MCP server. The server exposes tools like get the state of this sensor or create this automation. The AI decides which tools to call to answer your question, and shows you the result. Any client can talk to any server, which is what makes it a standard rather than a feature of one company.

Two MCP servers for Home Assistant

Home Assistant ships its own MCP server integration. It is deliberately narrow: clients can only see and control entities that have been exposed to Assist, and they cannot touch automations, dashboards or configuration.

ha-mcp is the community alternative, described by its maintainers as the unofficial and awesome Home Assistant MCP server. It is open source under the MIT licence, maintained by the homeassistant-ai group, and exposes roughly 87 tools: device control, entity states, history and statistics, automations, scripts, scenes, dashboards, backups, add-ons, the device registry and automation execution traces for debugging. It can read the whole home and, if you allow it, change it.

What it is like to use

In practice the reading side is the immediately useful part. Ask why the heating ran all night and the assistant pulls the thermostat history, the window sensors and the automation trace, then explains it. Ask which devices dropped off the Thread network this week and it lists them. The writing side, creating automations from a sentence, works and saves time, but early reviewers were right that it needs supervision. An assistant that can edit your home should be treated like a new employee with admin rights: useful, and worth a few rules.

What leaves your network

The server runs inside Home Assistant on the hub in your house. ha-mcp itself sends no telemetry. What does leave is the conversation: your question and the data the tools return for it go to the AI provider you chose, exactly as with any chat. When you are not asking anything, nothing is sent. If that trade-off is not acceptable, a local model can be used as the client, at the cost of capability.

The guardrails we set

• Read-only mode until you decide otherwise. The assistant can explain the home before it is allowed to change it.
• Per-tool permissions, so the tools that edit files or YAML stay off unless you turn them on.
• Automatic backups before any change to an automation, script or dashboard, which ha-mcp supports.
• Local-only access by default. Remote access from outside the house goes through an encrypted link that we set up only if you ask.
• A written record of what the assistant can and cannot touch, in the same handover pack as the rest of the system.

How it is installed

The recommended route is the ha-mcp custom component installed through HACS, which runs the server inside Home Assistant on every installation type without managing tokens. Docker and a Home Assistant add-on are the alternatives. A setup wizard on the project's site generates the configuration for Claude Desktop, Claude Code, ChatGPT, Gemini CLI, Cursor, VS Code and more than a dozen other clients. We do this as part of an install and add it to existing Home Assistant homes across Greater Montreal.