Home Assistant MCP — How Claude Builds Your Whole Smart Home With Zero YAML

Home Assistant is famously powerful, but its Automations, templates, and YAML pile up into a wall that scares off beginners. An MCP (Model Context Protocol) server makes the AI scale that wall. It's an interface that lets the AI see and *understand* your Home Assistant entities and services, not just send messages at them.

The creator's example lands hard. He asked Claude (in German) to build a baby dashboard with the Owlet sock and related sensors, organised into meaningful sections — and Claude built it. Not one line of YAML, no Automation editor, no template tinkering. His wife now uses the same setup herself to have Claude build Home Assistant automations. A household-level paradigm shift, condensed into one demo.

Security is sane by default. The first time the AI tries to call a new MCP tool, it asks for permission, so nothing runs silently behind your back. A quick test — 'how many entities does my system have?' — pulled a system-overview call from the Home Assistant Test MCP server and returned a detailed answer. The creator suggests naming MCP servers by environment (test vs prod) so the AI can keep them straight.

Claude Desktop users are basically done after the install. The video also covers Cursor, Antigravity, VS Code, WezTerm, and LM Studio. Most of these clients have an mcp.json somewhere in settings where you paste the Home Assistant MCP server URL block. LM Studio is the odd one out — its mcp.json lives under Developer → Server.

The recommended path is to start with one or two simple examples. Get a feel for how the AI interacts, what kind of code or automations it proposes, whether the quality holds up. Then escalate to real requirements — 'here's the requirement, I don't know how to implement it, build me a dashboard' — and watch it produce something workable. One or two follow-up prompts and it runs. The creator calls this a paradigm shift, not a toy.

For privacy-minded users, Ollama and LM Studio let you run everything locally. For a system that controls your home, the local-first stance makes obvious sense. The honest trade-off: local models usually don't match the quality of the big US cloud models. Whether to lean cloud or local is a values call, and the creator names that trade-off without flinching.