Rank #04 · Daily · Agent-Skill
HA-MCP: MCP Server to Configure and Automate Home Assistant with AI
Point an AI assistant at your Home Assistant and ask it to debug automations, edit dashboards, or manage devices — through a standalone MCP server, not just Assist-exposed entities.
What it does
HA-MCP is a Model Context Protocol server for configuring and automating Home Assistant. GitHub description: "Configure and automate Home Assistant with Claude, ChatGPT, and other AI assistants." Docs: homeassistant-ai.github.io/ha-mcp. Built on FastMCP. License: MIT.
From today's README (as verified):
- Ships a large tool catalog (README: 87 tools) for fuzzy entity search, service/device control, automations, scripts, helpers, dashboards, areas, zones, groups, calendars, blueprints, history, traces, backups, apps, and the device/entity registry.
- Safety controls: a Read Only Mode toggle, per-tool enable/disable, user-approval security policies, and automatic edit backups.
- Recommended install is the in-process HA-MCP integration. It runs on Home Assistant OS, Supervised, Container, and Core, and needs Home Assistant 2026.8.0 or newer. On 2026.8–2026.10, README points you to HACS. On Home Assistant 2026.11 or newer, README says the preferred path is Settings → Marketplace, with HA-MCP in the default catalog (no custom GitHub repository required). TidyHow cites that install path from the project README; we are not independently dating the 2026.11 Marketplace launch or treating it as the proven cause of today's star surge.
- Setup asks for an administrator long-lived token, or a dedicated local-only administrator account. Clients connect via a Home Assistant webhook URL and/or a local network port. Remote access starts disabled.
- Other runtimes exist (Home Assistant app/add-on, Docker, PyPI/
uvxHTTP, local stdio). README warns: configure exactly one install method per client. File and YAML editing tools are opt-in beta and need a second integration entry. - A Setup Wizard covers 15+ clients (Claude Desktop / Claude Code, ChatGPT, Gemini CLI, VS Code, Cursor, and others). A hosted demo server is documented for a trial without your own instance.
Who it's for
- Home Assistant users who want an AI client to build and debug automations, dashboards, and helpers — not only toggle already-exposed lights.
- People running Home Assistant 2026.8+ who can grant an administrator token (or a dedicated admin account) to a local MCP server.
- Teams wiring Claude, ChatGPT, or other MCP clients into a self-hosted smart home.
Less ideal if you have no Home Assistant instance, only need Assist-style control of already-exposed entities (the built-in MCP Server may suffice), or are unwilling to give an MCP server administrator-level access.
Stars / language / license
7,372 stars · Python · MIT
Data verified: 2026-11-12 09:00 Asia/Shanghai (GitHub API get_repository + README)
Trending note: +645 stars today (captured 08:58 Asia/Shanghai)
Category tags
Agent-Skill · MCP · Home-Assistant · Python · FastMCP · home-automation · Claude · ChatGPT
How it differs from similar projects
Compared with Home Assistant's built-in MCP Server. The official integration rides the Assist pipeline and is limited to entities you have exposed to Assist, plus the intents Assist already understands. HA-MCP's README positions itself as a standalone server for configuring, building, and debugging the whole instance — automations, dashboards, traces, backups, HACS/apps, and the registries. Use the built-in server for voice-style control of exposed devices; use HA-MCP when the assistant also needs to maintain the setup.
Compared with hand-rolled REST / WebSocket scripts. HA-MCP packages a documented MCP tool surface, client wizard, and in-process Home Assistant install. You still have to operate the server and its credentials.
Compared with cloud home-automation assistants. HA-MCP talks to your Home Assistant. Remote webhook access is optional and, as the README notes, a leaked administrator token works from anywhere — prefer a dedicated local-only admin account.
Caveats. Administrator credentials are powerful. File/YAML tools and some dashboard/YAML helpers are beta. Tool count and Marketplace wording are project-README claims. No independent reliability benchmarks are claimed here.