Home Assistant MCP Server

A Model Context Protocol (MCP) server that lets AI assistants control Home Assistant. Built with Python and FastMCP.

Works with Claude, GPT-4, Cursor, Kiro, and any MCP-compatible client.

What it does

• 40 tools covering lights, climate, covers, locks, media players, vacuums, fans, cameras, alarms, and more
• BM25 tool search — LLMs discover tools on demand instead of receiving all 40 schemas upfront
• MCP Resources for read-only entity, area, device, and service data
• MCP Prompts for guided workflows (automation creation, troubleshooting, energy optimization)
• Tool annotations — readOnlyHint lets clients skip confirmation prompts for safe operations
• Smart error log parsing — deduplicates and summarises HA error logs instead of dumping raw text
• Context-aware responses — compact list responses, domain-filtered queries, progress reporting
• Async throughout with TTL-based caching to reduce API load

Installation

uvx (recommended)

No install needed. Just configure your MCP client:

{
  "mcpServers": {
    "homeassistant": {
      "command": "uvx",
      "args": ["homeassistant-mcp"],
      "env": {
        "HASS_HOST": "http://homeassistant.local:8123",
        "HASS_TOKEN": "your_long_lived_access_token_here"
      }
    }
  }
}

pip

pip install homeassistant-mcp
homeassistant-mcp

From source

git clone https://github.com/robbrad/homeassistant-mcp.git
cd homeassistant-mcp
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -e ".[dev]"
homeassistant-mcp

Configuration

The server needs two environment variables:

Optional:

These can be set via environment variables (as shown in the MCP client configs above) or in a .env file. See .env.example.

MCP Client Setup

Claude Desktop

Add to your config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "homeassistant": {
      "command": "uvx",
      "args": ["homeassistant-mcp"],
      "env": {
        "HASS_HOST": "http://homeassistant.local:8123",
        "HASS_TOKEN": "your_token_here"
      }
    }
  }
}

Cursor / Kiro / Other MCP Clients

Same config format. Place it in your client's MCP configuration file (e.g. .cursor/mcp.json, .kiro/settings/mcp.json).

The server uses stdio transport, which is the standard for MCP.

How Tool Discovery Works

With 40 tools, the server uses BM25 search to keep the LLM's context lean. Instead of sending all 40 tool schemas upfront, the LLM sees 8 tools:

The LLM calls discover_tools() to see what's available, then search_tools(query="lights brightness") to get the full schema, then calls the tool directly.

All Tools

Device Control (16 tools)

Automation & Scenes (3 tools)

Input Helpers (5 tools)

API & State (4 tools)

History (3 tools)

Specialized (5 tools)

General (4 tools)

MCP Resources

Development

git clone https://github.com/robbrad/homeassistant-mcp.git
cd homeassistant-mcp
python -m venv venv
source venv/bin/activate
pip install -e ".[dev]"

# Run tests
pytest

# Format and lint
black src/ tests/
ruff check src/ tests/

# Type check
mypy src/

Publishing

Commits to main auto-publish to PyPI via GitHub Actions + python-semantic-release:

• fix: commits → patch bump (3.1.0 → 3.1.1)
• feat: commits → minor bump (3.1.0 → 3.2.0)
• BREAKING CHANGE: → major bump (3.1.0 → 4.0.0)

Architecture

AI Assistant  <-->  MCP Server (FastMCP/stdio)  <-->  Home Assistant REST API
                         |
                   +-----+-----+
                   |           |
                 Tools      Cache
                 Layer      Layer

• FastMCP Server handles MCP protocol over stdio with BM25 tool search
• Home Assistant Client is an async httpx client with auth and connection pooling
• Cache Layer provides TTL-based caching for states and entities
• Tools Layer — 40 tools with annotations, tags, timeouts, and context logging
• Resources — 5 read-only data endpoints with compact response envelopes
• Prompts — 13 guided workflows for automation, troubleshooting, and optimization

License

MIT