garmin-mcp

An MCP server that exposes your Garmin Connect data to Claude as tools. Ask things like "how did I sleep last night?" or "summarise my training load this week" and Claude answers using your real Garmin data instead of you copy-pasting screenshots from the app.

Single-user, read-only. Two ways to run it:

Tools

Every response is a Pydantic model serialised to JSON, with null for fields Garmin did not record.

Quick start — Claude Desktop

Requires Python 3.12+ and uv (install with curl -LsSf https://astral.sh/uv/install.sh | sh).

1. Authorise once

uvx garmin-mcp login

Prompts for your Garmin email, password, and MFA code (if enabled), then saves session tokens to your user cache directory. You won't be prompted again until the tokens eventually expire (typically weeks to months).

2. Add the server to Claude Desktop

Edit claude_desktop_config.json:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Windows: %APPDATA%\Claude\claude_desktop_config.json

• Linux: ~/.config/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "garmin": {
      "command": "uvx",
      "args": ["garmin-mcp"]
    }
  }
}

Restart Claude Desktop. The Garmin tools appear in the tool picker. Ask Claude "what was my resting heart rate this week?" to test.

3. (Optional) Set credentials for unattended re-auth

By default, when Garmin tokens expire you'll see a "saved Garmin session is invalid" error and you'll need to re-run uvx garmin-mcp login. To skip that step, put your credentials in the config so the server can silently re-authenticate:

{
  "mcpServers": {
    "garmin": {
      "command": "uvx",
      "args": ["garmin-mcp"],
      "env": {
        "GARMIN_EMAIL": "you@example.com",
        "GARMIN_PASSWORD": "your-garmin-password"
      }
    }
  }
}

Anyone with read access to this file can see these credentials.

Where session tokens are stored

garmin-mcp login writes session tokens to your platform's user cache directory:

Delete the garth/ directory to "log out" of Garmin.

Self-hosted HTTP (Claude.ai web/mobile)

If you want the connector available from Claude.ai on the web or your phone, run the same server in HTTP mode. The serve subcommand wraps it in an OAuth 2.1 layer with PKCE and Dynamic Client Registration so Claude.ai can connect to it as a custom connector.

See DEPLOY.md for the Cloud Run walkthrough. The short version:

docker build -t garmin-mcp .
docker run --rm -p 8080:8080 \
  -e MCP_ISSUER_URL=http://localhost:8080 \
  -e MCP_AUTH_PASSWORD=$(openssl rand -base64 24) \
  -e JWT_SECRET=$(openssl rand -base64 48) \
  -e GARMIN_EMAIL=you@example.com \
  -e GARMIN_PASSWORD=your-garmin-password \
  garmin-mcp

For Cloud Run, the always-free tier covers personal usage. Expect under $1/month.

How auth works (HTTP mode)

The server is its own OAuth 2.1 authorisation server. When you add the connector in Claude.ai, Claude registers itself using RFC 7591 Dynamic Client Registration, then sends you through a PKCE-protected flow. You enter the password set as MCP_AUTH_PASSWORD, and the server issues a 24-hour JWT access token plus a refresh token that rotates on every use.

This is intentionally minimal: one password, one user. Anyone with the password can read your Garmin data.

Data availability

Garmin returns sparse data depending on which watch you wear, how long you've worn it, and what features your model supports. Every tool follows the same convention: when a field isn't recorded, the response carries null for that field (and often a note explaining the absence) rather than erroring.

A few specific cases worth knowing about:

• get_training_load.current_status = "NO_STATUS_2" and get_hrv_status.status = "NONE" mean Garmin doesn't have enough recent activity history to compute the metric. They fill in naturally after ~7 consecutive days of sustained activity or watch wear.

• VO2 max only updates after qualifying activities (runs, rides). get_fitness_metrics walks back up to 7 days to surface your most recent reading rather than returning null on a rest day.

• get_stress zone-minute breakdown (rest_minutes, low_minutes, etc.) can come back null on partial-data days even though avg_stress and the timeline are populated.

• HRV, training readiness, endurance score, hill score, and fitness age all require a recent compatible watch (Fenix 6+ / Forerunner 245+ / similar). Older watches simply won't report them.

If a tool seems to return less than you'd expect, check the same metric in the Garmin Connect app or on connect.garmin.com for the same date. If Garmin shows it there and we return null, that's a parser bug — file an issue with the date and the field name and we can usually map it in a follow-up release.

Security caveats

• This is single-user software. Don't run it as a shared service for multiple Garmin accounts — you'd be holding other people's credentials, and it likely violates Garmin's ToS.

• Garmin credentials and session tokens live on your local machine. Treat any password you put in a JSON config file as compromised in the long term — use a dedicated Garmin account if that's a concern.

• The unofficial garminconnect library can break when Garmin changes their internal API. If a tool starts returning empty data, check that package's changelog.

• In HTTP mode, registered DCR clients and refresh tokens live in process memory and disappear on restart. Access tokens (JWTs) survive because they are stateless.

• This server is read-only. It does not write activities, edit profile fields, or upload anything to Garmin.

Project layout

garmin-mcp/
├── pyproject.toml
├── Dockerfile
├── README.md
├── DEPLOY.md
└── src/
    └── garmin_mcp/
        ├── __init__.py
        ├── __main__.py        # python -m garmin_mcp -> CLI
        ├── cli.py             # argparse entry: stdio / serve / login
        ├── server.py          # FastMCP app, tools, login UI
        ├── garmin_client.py   # garminconnect wrapper
        ├── auth.py            # OAuth 2.1 provider
        ├── cache.py           # TTL cache
        ├── paths.py           # token directory resolution
        └── models.py          # Pydantic response models

Contributing

git clone https://github.com/Tyler-Irving/garmin-mcp.git
cd garmin-mcp
uv sync --extra dev

uv run garmin-mcp login                       # one-time interactive login
uv run mcp dev src/garmin_mcp/server.py       # inspect tools in MCP Inspector
uv run garmin-mcp                             # stdio mode
uv run garmin-mcp serve                       # HTTP mode

uv run pytest                                 # tests
uv run ruff check . && uv run ruff format --check .
uv run mypy src tests

Acknowledgements

• garminconnect by cyberjunky for doing the hard work of reverse-engineering the Garmin Connect API.

• The Model Context Protocol team for the SDK.