mcp-garmin-connect 0.3.1

LLM-agnostic Garmin Connect MCP server for training, recovery, sleep, stress, VO2 Max, race predictions, and running summaries.

Use the same local Garmin tools from any MCP client — Claude Desktop, Claude Code, Cursor, Continue, Hermes, opencode — or from provider-backed demo agents using DeepSeek, OpenAI, Claude, OpenRouter, or Gemini.

See It In Action

Recorded with garmin-mcp serve --demo, which serves deterministic fictional sample data — no Garmin credentials and no real health data involved:

uvx mcp-garmin-connect serve --demo

Works With Any MCP Harness

This server implements standard MCP over stdio and Streamable HTTP, so there is no harness-specific adapter to maintain. If your tool speaks MCP, it works:

Any client that supports MCP over stdio or Streamable HTTP can use every tool and prompt in this server with no extra code.

Why This Exists

Most fitness advice from LLMs is generic. This server lets an LLM inspect your real Garmin Connect data first, then answer with context from your activities, recovery markers, HRV, sleep, stress, VO2 Max, and running history.

The core server is provider-neutral. DeepSeek, OpenAI, Claude, OpenRouter, and Gemini integrations are optional demo bridges for clients that do not speak MCP directly.

Two Garmin Backends

GARMIN_AUTH_MODE picks the strategy:

• auto (default) — official Personal API when GARMIN_PERSONAL_API_KEY is set, else email/password.
• personal / unofficial — force one backend.
• hybrid — run both: wellness and raw-file reads go to the official API, Garmin-derived metrics go to email/password, with automatic fallback when a call fails. This is the only mode that returns derived metrics and raw .FIT files from one process. It needs both credential sets. Tools that need a metric the active backend cannot see return that field as null with an errors entry explaining why, instead of failing the whole call. Setup and endpoint mapping: docs/personal-api.md.

Features

• Published on PyPI — run it with uvx mcp-garmin-connect, no clone or manual install.
• 24 Garmin tools and 4 prompt templates for training and recovery analysis.
• Two Garmin backends, switchable or combined — the official Connect Personal API (self-serve key, stable) and the unofficial email/password flow (wider metric coverage). GARMIN_AUTH_MODE=auto picks one; hybrid runs both and routes each call to the backend that owns the data.
• Raw activity files — download_activity_file fetches the original .FIT (ZIP) through the Personal API.
• Demo mode — serve --demo runs on fictional sample data with no credentials, for try-outs and recordings.
• Local MCP server over stdio or Streamable HTTP (/mcp).
• Endpoint-level TTL cache with single-flight factories — concurrent tool calls share one Garmin request instead of stampeding the API.
• Global request concurrency cap so parallel snapshots stay within a safe request rate.
• Partial-error reporting — composite tools include an errors array when an endpoint fails, instead of silently returning nulls.
• Provider demo bridge for deepseek, openai, claude, openrouter, and gemini.
• Explicit opt-in before Garmin health/activity data is sent to any external LLM provider.
• Mocked test suite plus optional live Garmin smoke tests.

Quick Start

Option A — published package (no clone needed)

uvx mcp-garmin-connect login
uvx mcp-garmin-connect serve --transport stdio

uvx runs the latest PyPI release in an isolated environment. Export GARMIN_EMAIL / GARMIN_PASSWORD (or keep a .env in the directory you launch it from) before the first login.

Option B — from source

git clone https://github.com/kavakoss/mcp-garmin-connect
cd mcp-garmin-connect
uv sync
cp .env.example .env

Pick a backend in .env:

# Official Personal API (recommended): create a key at
# https://connect.garmin.com/app/settings/personalApiToken
GARMIN_AUTH_MODE=personal
GARMIN_PERSONAL_API_KEY=your-key-here

or the unofficial email/password flow:

GARMIN_AUTH_MODE=unofficial
GARMIN_EMAIL=you@example.com
GARMIN_PASSWORD=your-password

The unofficial flow needs a one-time interactive login (MFA supported); the Personal API needs no login step:

uv run garmin-mcp login                  # unofficial backend only
uv run garmin-mcp personal-api --check   # official backend: token + sample per endpoint
uv run garmin-mcp doctor --live

List available MCP tools, then run the server:

uv run garmin-mcp tools
uv run garmin-mcp serve --transport stdio

HTTP transport:

uv run garmin-mcp serve --transport http --host 127.0.0.1 --port 8765
# clients connect to http://127.0.0.1:8765/mcp

Windows (PowerShell) equivalents: uv sync, Copy-Item .env.example .env, notepad .env, then the same uv run garmin-mcp ... commands.

MCP Client Setup

Generic stdio config using the published package (uvx):

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

From a local checkout, replace command/args with:

{
  "mcpServers": {
    "garmin": {
      "command": "uv",
      "args": [
        "--directory",
        "/path/to/mcp-garmin-connect",
        "run",
        "garmin-mcp",
        "serve",
        "--transport",
        "stdio"
      ]
    }
  }
}

With the local checkout, credentials are also read from the repo's .env file, so the env block is optional.

opencode (opencode.json):

{
  "mcp": {
    "garmin": {
      "type": "local",
      "command": [
        "uv",
        "--directory",
        "/path/to/mcp-garmin-connect",
        "run",
        "garmin-mcp",
        "serve",
        "--transport",
        "stdio"
      ],
      "enabled": true
    }
  }
}

Per-client examples (Claude Desktop, Claude Code, Cursor, Continue, Hermes, generic stdio/HTTP) live in docs/clients.

MCP Tools

Activity windows

get_recent_activities accepts either a lookback (days) or an explicit window:

{ "days": 30 }                                          // newest 30 days
{ "start_date": "2026-08-01", "end_date": "2026-08-31" } // August 2026 only

Both dates are inclusive and YYYY-MM-DD; end_date defaults to today. Explicit dates win over days and make a past window queryable without pulling every newer activity first — the pager walks backwards and stops once it reaches start_date. The response then includes the resolved window so callers can see exactly what was queried. Invalid or reversed dates return an error field instead of silently returning everything.

MCP Prompts

• recovery_check
• weekly_training_review
• activity_analysis
• race_plan_context

Configuration

All settings come from environment variables (.env is loaded automatically).

Provider bridge variables (optional): DEEPSEEK_API_KEY, OPENAI_API_KEY, ANTHROPIC_API_KEY, OPENROUTER_API_KEY, GEMINI_API_KEY, plus the matching *_BASE_URL and *_MODEL overrides.

Field units

Garmin's raw payloads mix units, and some fields are named after a unit that Garmin does not actually use (see issue #3: recoveryTime is minutes, not hours). This table is the contract — if you add a field, add its row.

Daily arrays (daily on the sleep, stress, RHR, HRV and body-battery tools) are newest-first — index 0 is the most recent day. Sort client-side if you need chronological order.

recovery_time_hours is only populated when the watch/account provides Training Readiness — a Forerunner 165, for example, returns no readiness data, so the whole training_readiness object is null. Other tools that lean on it degrade the same way rather than inventing numbers.

How It Stays Fast And Quiet

• Endpoint-level cache. Sleeping, stress, RHR, and activity fetches are cached per day/limit and shared across every tool that needs them — a get_full_snapshot burst reuses the same responses instead of refetching.
• Single-flight. If two tool calls ask for the same missing key, only one hits Garmin; the other waits and reuses the result.
• Failures are never cached. A failed endpoint is retried on the next call.
• Concurrency cap. All outbound Garmin calls pass through one bounded semaphore (GARMIN_MAX_CONCURRENCY) so parallel sections cannot flood the API.
• Visible partial failures. When an optional endpoint fails, composite tools add an errors array (source + exception) to their output.

Provider Demo Agents

Provider API access may require a paid account or credits even if the provider's web app has a free tier.

DEEPSEEK_API_KEY=sk-...
OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...
OPENROUTER_API_KEY=sk-or-...
GEMINI_API_KEY=...

uv run garmin-mcp ask "How is my recovery today?" --provider deepseek --allow-external-health-data
uv run garmin-mcp ask "Summarize my running for 90 days." --provider openai --allow-external-health-data
uv run garmin-mcp ask "What should I focus on this week?" --provider claude --allow-external-health-data

The --allow-external-health-data flag is required because Garmin tool results can be sent to the selected provider.

Privacy And Safety

• Your .env file and Garmin token cache stay local and are ignored by git.
• MCP clients that run locally can call Garmin tools without sending provider API keys to this project.
• Provider demo commands such as garmin-mcp ask can send Garmin health/activity results to the selected LLM provider, so the CLI requires --allow-external-health-data.
• get_zones returns only zone/threshold fields from the Garmin profile, not the raw profile (gender, birth date, weight, and similar PII are trimmed).

Testing

uv run garmin-mcp doctor
uv run garmin-mcp doctor --live
uv run garmin-mcp tools
uv run ruff check .
uv run pytest

CI runs on Linux (Python 3.11, 3.12, 3.13) and Windows (3.12). Optional live tests are gated and never run in CI:

$env:GARMIN_LIVE_TEST = "1"
uv run pytest tests/live

Troubleshooting

• Garmin rate-limited the request — lower GARMIN_MAX_CONCURRENCY (for example to 3) and retry after a few minutes.
• Garmin authentication failed — check GARMIN_EMAIL/GARMIN_PASSWORD, then run uv run garmin-mcp login from a real terminal (MFA prompts need interactive input).
• Garmin rejected the Personal API key — keys expire after 30/90/180/365 days. Create a new one at https://connect.garmin.com/app/settings/personalApiToken, update GARMIN_PERSONAL_API_KEY, then uv run garmin-mcp personal-api --check.
• personal api required — that tool only exists on the official backend. Set GARMIN_PERSONAL_API_KEY and GARMIN_AUTH_MODE=personal (or hybrid).
• A derived metric is null on the personal backend — training status/load, readiness, VO2 Max, race predictions and personal records are not in Garmin's Personal API. Switch to GARMIN_AUTH_MODE=unofficial when you need them.
• A tool returns errors — the array names the endpoint that failed; the rest of the payload is still valid partial data.

Garmin Disclaimer

This project uses the community garminconnect Python package and Garmin Connect endpoints. Garmin Connect is not a public API for individual open-source projects, and authentication or endpoint behavior may change without notice.

Some metrics depend on Garmin device capabilities. For example, a device may expose VO2 Max and Training Effect but not native Garmin Training Load or Training Status. Zone availability depends on what the Garmin user settings endpoint exposes for your account.

License

This project is licensed under the MIT License.

Made by kavakoss. Inspired by Jack-Abyss/claude-garmin, with a clean-room implementation focused on LLM-agnostic MCP usage.