Garmin Personal MCP

Garmin Personal MCP is a free, open-source MCP server that connects your Garmin Connect account to Claude, Cursor, VS Code, or any other MCP client. Ask about last night's sleep, HRV trends, Body Battery, stress, training load, weight, or a specific run, all from your real Garmin account. It runs on Garmin's official Personal API with a revocable key, so it never needs your Garmin password, and it's read-only.

Published on PyPI as garmin-personal-mcp and in the official MCP Registry as io.github.Mason-Levyy/garmin-personal-mcp.

Unofficial project, official API. This isn't made by Garmin, but it only talks to Garmin's documented Personal API through Garmin's own Python client.

How is this different from other Garmin MCP servers?

Most Garmin MCP servers sign in with your Garmin password through the unofficial garminconnect library, which breaks whenever Garmin changes its login. In October 2026 Garmin opened a Personal API for individuals. This server uses it:

• No password, no scraping: a revocable Personal API key you create in Garmin Connect, exchanged for short-lived tokens
• The whole Personal API: every working endpoint in Garmin's reference has a tool, with every documented filter
• Any date range: ranges longer than an endpoint allows (28 days for steps, 31 for sleep, and so on) are split into requests for you
• Model-sized responses: all-day sample arrays are left out unless you ask for them
• Read-only: the Personal API can't change your account, and neither can this server

Quickstart

1. Save your Personal API key (one-time; see Authentication):

   bash

   uvx garmin-personal-mcp auth

2. Add the server to your client. The install buttons at the top do this in one click for Cursor and VS Code.

   Claude Code

   bash

   claude mcp add garmin -- uvx garmin-personal-mcp

   Or install it as a plugin, which asks for your key in place of step 1:

   /plugin install garmin-personal-mcp --marketplace Mason-Levyy/garmin-personal-mcp

   Claude Desktop (claude_desktop_config.json), Cursor (~/.cursor/mcp.json), Windsurf (~/.codeium/windsurf/mcp_config.json), and Cline (cline_mcp_settings.json) all take the same entry:

   json

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

   VS Code, Codex CLI, Gemini CLI

   VS Code (or use the button above):

   bash

   code --add-mcp '{"name":"garmin","command":"uvx","args":["garmin-personal-mcp"]}'

   Codex CLI:

   bash

   codex mcp add garmin -- uvx garmin-personal-mcp

   Gemini CLI:

   bash

   gemini mcp add garmin uvx garmin-personal-mcp

3. Ask it something: "how did I sleep last night?", "is my HRV trending down this month?", "compare my last three long runs".

Requires uv. Any MCP client that speaks stdio or streamable HTTP works, not just Claude.

Finding this useful? A ⭐ on GitHub helps other Garmin users find it.

Authentication

1. Sign in to Garmin Connect on the web and open Settings → Personal API.
2. Create a key and choose how long it lasts (30 days to a year). Garmin shows it once.
3. Run garmin-personal-mcp auth and paste it at the hidden prompt. The server checks it with Garmin before saving it to ~/.config/garmin/connect-personal-api-key with owner-only permissions.

Where there's no terminal to prompt in (a script, or a shell an agent runs), pipe the key in instead, e.g. pbpaste | garmin-personal-mcp auth.

garmin-personal-mcp auth --check reports whether the saved key still works, without prompting. Garmin emails you about a week before a key expires: create a new one and run auth again. No restart is needed; the server reads the key on every call.

Instead of the key file you can set GARMIN_PERSONAL_API_KEY, which takes precedence. The key file and bearer-token cache are the same ones Garmin's gc-personal CLI uses, so a key saved by either works for both.

Docker Compose

Run the server as a background HTTP service using the prebuilt image. Save this as compose.yaml:

services:  garmin:    image: ghcr.io/mason-levyy/garmin-personal-mcp:latest    command: ["--http", "--host", "0.0.0.0"]    environment:      - GARMIN_PERSONAL_API_KEY    ports:      - "127.0.0.1:8491:8491"    volumes:      - garmin-config:/home/garmin/.config/garmin      - ./downloads:/home/garmin/Downloads/garmin    restart: unless-stopped
volumes:  garmin-config:

latest is the newest release, the same version PyPI serves. To stay on one version, use a version tag such as :0.2.0 instead.

From the directory containing compose.yaml, create the downloads folder and pull the image:

mkdir -p downloadsdocker compose pull

Activity files from garmin_download_activity land in ./downloads. The tool reports the path inside the container (/home/garmin/Downloads/garmin/...), which is the same file. The container runs as uid 1000, so on Linux, if your uid isn't 1000, run sudo chown 1000 downloads.

Don't mount an existing ~/.config/garmin from the host into the container. The key file there is readable only by its owner, so the container can't read it unless your uid is 1000. Use one of the options below instead.

The container paths above and uid 1000 stay the same across releases.

Authentication

Choose one of these authentication options:

1. Add your API key to a .env file

Create .env next to compose.yaml and add your Personal API key:

GARMIN_PERSONAL_API_KEY='your-personal-api-key'

Compose loads this file automatically. You do not need to run auth.

2. Use the interactive auth command

Leave GARMIN_PERSONAL_API_KEY unset in your shell and .env, then run:

docker compose run --rm garmin auth

Paste your Personal API key at the prompt. It is saved for subsequent starts.

Start the server

docker compose up -d

In your MCP client, add a Streamable HTTP server with the URL http://127.0.0.1:8491/mcp.

The server only accepts requests addressed to 127.0.0.1 or localhost, and it has no authentication. To reach it from another machine, keep the port on 127.0.0.1, put an authenticating proxy or tailnet in front (see Remote / HTTP mode), and add the hostname clients use to command:

    command: ["--http", "--host", "0.0.0.0", "--allowed-host", "garmin.example.ts.net"]

Without --allowed-host, requests to any other hostname are rejected.

Build from source

Clone the repository and use compose.yaml in its root, which builds the image from source. From the repository directory, run:

docker compose build

Then follow the authentication and startup steps above. After updating the source, rebuild and restart with:

docker compose up --build -d

Tools

All tools take dates as YYYY-MM-DD. Range tools default to the 7 days ending today. Tools that return a list wrap it as {"results": [...]}.

Weights are in grams, distances in meters, and durations in seconds, as Garmin returns them.

Remote / HTTP mode

The default transport is stdio. For network clients:

garmin-personal-mcp --http --host 127.0.0.1 --port 8491

This serves streamable HTTP at /mcp. There is no built-in authentication. Never expose it to the internet. Bind to localhost and front it with something that authenticates for you: a VPN or tailnet (e.g. tailscale serve), an authenticating reverse proxy, or an OAuth-aware MCP gateway. Behind a proxy, pass --allowed-host <proxy hostname> (repeatable) so the DNS-rebinding check accepts its Host header.

Configuration

Troubleshooting

• "Garmin rejected the Personal API key": the key expired or was deleted. Create a new one and run garmin-personal-mcp auth.
• "No usable Garmin Personal API key was found": run garmin-personal-mcp auth, or set GARMIN_PERSONAL_API_KEY in the client config.
• {"no_data": ...}: Garmin has nothing for that date yet (today's HRV before you've slept, for example), or your device doesn't record that metric.
• HTTP 400 on a long range: an endpoint without a documented limit rejected the range. Ask for a shorter one.

FAQ

Is this the official Garmin API? Yes. It uses Garmin's Personal API, which anyone with a Garmin Connect account can enable without applying, through Garmin's Apache-licensed Python client. The project itself is independent.

Can it change my Garmin data? No. The Personal API is read-only. The one tool that writes anything, garmin_download_activity, saves a file on your own machine.

Which MCP clients does it work with? Any client that runs a stdio or streamable HTTP server, including Claude Desktop, Claude Code, Cursor, Windsurf, Cline, and VS Code. AI coding agents can follow llms-install.md.

Where does my data go? The server talks only to Garmin. Your key and bearer token stay in ~/.config/garmin/. Tool results go to whichever MCP client and model you connect it to.

Can I use it for other people's accounts? No. Garmin's Personal API is for your own data. Apps serving other users need Garmin's business developer program.

How it works

• garmin-personal-mcp auth saves the key; each call exchanges it for a bearer token through Garmin's PersonalApiAuthenticator, which caches the token until an hour before it expires. A token rejected early gets one retry with a fresh exchange.
• Each tool maps to one endpoint on personal.connectapi.garmin.com (garmin_get_day combines several). Ranges longer than an endpoint's documented maximum are fetched in consecutive windows and concatenated.
• Bulky per-minute sample arrays are dropped by default to keep responses readable for a model; include_samples or chunks bring them back.

Development

git clone https://github.com/Mason-Levyy/garmin-personal-mcpcd garmin-personal-mcpuv syncuv run pytest

Tests run against synthetic fixtures; no account needed. Lint and formatting are enforced with ruff (uv run ruff check ., uv run ruff format .).

Contributing

Issues and pull requests are welcome, especially when Garmin adds endpoints to the Personal API. See CONTRIBUTING.md for setup, style, and the PR checklist, and SECURITY.md for how to report vulnerabilities privately. Release history is in CHANGELOG.md.

Support the project

If this server saves you time, please star the repo. Stars are how other Garmin users find it in GitHub search and MCP directories. Bug reports and endpoint requests are just as welcome in issues.

License

MIT. Garmin and Garmin Connect are trademarks of Garmin Ltd. or its subsidiaries; this project is not affiliated with or endorsed by Garmin.