garmin-mcp-gt 0.1.3

Garmin Connect MCP server — exposes health and activity data as tools for Claude

Fuel

A calorie and activity tracker for cyclists, powered by Claude and Garmin Connect.

Log food in plain English, get a calorie balance against your activity, and ask questions about your fitness data — all from a mobile-friendly web app.

Features

• Food logging — describe what you ate in plain English; Claude estimates the calories
• Activity balance — live calorie in vs. burned, updated from Garmin
• Conversational health advisor — ask "how were my steps today?" or "how did I sleep last night?" and get answers using your real Garmin data
• Push nudges — scheduled notifications with time-aware advice (morning vs. evening)
• Multi-user — each user has their own data, secured with Firebase Auth

Using the app

The hosted app is at https://pike-477416.web.app

1. Sign in with Google or create an email/password account
2. Log food in the text box ("had porridge and a coffee")
3. Connect Garmin to enable activity data and the Ask tab (see below)
4. Tap Ask to chat with Claude about your health data

Connecting Garmin

Garmin uses session tokens that must be generated on your own machine (there is no public OAuth API). Setup takes about two minutes:

1. Install the CLI

pip install garmin-mcp-gt

2. Authenticate with Garmin

garmin-setup

Enter your Garmin Connect email and password when prompted. Tokens are saved to ~/.garmin_tokens/.

3. Generate an upload token in the app

In the Fuel app, scroll to the bottom of the Log tab and tap Connect Garmin. This generates a 15-minute upload token.

4. Upload your tokens

garmin-upload-tokens --token "paste-token-here"

That's it. Your Garmin data is now available in the app and the Ask tab will use it.

Tokens expire periodically (weeks to months). Re-run garmin-setup then garmin-upload-tokens when needed.

Using with Claude Code / Claude Desktop

garmin-mcp-gt is also an MCP server for use in Claude Code sessions. After pip install garmin-mcp-gt and garmin-setup, add to ~/.claude/settings.json:

{
  "mcpServers": {
    "garmin": {
      "command": "garmin-mcp-gt"
    }
  }
}

Then ask Claude things like "am I sedentary today?" or "how does my sleep affect my training load?".

Architecture

• Frontend — React PWA, hosted on Firebase Hosting
• Backend — FastAPI on Cloud Run (europe-west1)
• Database — SQLite with Litestream replication to GCS (zero managed DB cost)
• Garmin tokens — encrypted at rest with Cloud KMS, per-user in SQLite
• Auth — Firebase Auth (Google Sign-In + email/password)
• AI — Claude API (food parsing, recommendations, conversational tools)

Self-hosting

The backend is a standard Cloud Run service. To deploy your own instance:

1. Fork this repo
2. Create a GCP project and enable the required APIs
3. Run backend/setup-wif.ps1 to configure GitHub Actions auth
4. Run backend/setup-secrets.ps1 to store your secrets
5. Add GitHub Actions secrets (see backend/setup-wif.ps1 output)
6. Push to main — CI deploys everything automatically

Infrastructure is managed with OpenTofu — see terraform/.

Development

# Python tests
python -m pytest

# Frontend tests
cd ui && npm test

# Run backend locally
DB_PATH=./dev.db uvicorn backend.main:app --reload