mcp-garmin
Exposes Garmin Connect data through MCP, enabling retrieval of activities, health metrics, training status, and workouts, with OAuth authentication and confirmed write operations.
README
mcp-garmin
An MCP server that exposes Garmin Connect over Streamable HTTP, authenticated with GitHub OAuth.
It is built for remote use — as a custom connector in claude.ai / Claude
Desktop, or over claude mcp add --transport http — rather than as a local
stdio server. Two consequences follow from that:
- Clients authenticate with OAuth, handled by FastMCP's
GitHubProvider. The server publishes/.well-known/oauth-authorization-serverand supports Dynamic Client Registration, which is what claude.ai requires. - Garmin credentials never reach the client. The interactive sign-in runs once; the resulting OAuth tokens live on a mounted volume and refresh themselves.
Data access is provided by garminconnect.
Why a GitHub OAuth app?
It is reasonable to ask what GitHub has to do with reading your own Garmin data. Nothing — GitHub is only the login screen. The chain that puts it there:
- claude.ai will not accept a static token. Its connector dialog can send fixed request headers, but that feature is in limited beta. Without it the only way to add a remote MCP server is OAuth.
- OAuth needs something that authenticates a human. The server must run an authorization flow, which means a browser sign-in that proves who you are.
- The server has no user database, and should not have one. FastMCP's
OAuthProxytranslates between the MCP client and an existing identity provider; it does not store users or passwords itself.
So an identity provider is required, and GitHub is a convenient one. The app
requests the user scope only — it reads your profile, not your repositories —
and the single field taken from it is your login, which is compared against
ALLOWED_GITHUB_LOGINS.
Nothing depends on GitHub specifically. FastMCP ships providers for Google,
Azure, Auth0, Keycloak, Discord, WorkOS and others; swapping is a one-line
import change in server.py plus the matching credentials. What will not
work is FastMCP's InMemoryOAuthProvider — it simulates the flow for tests and
authenticates nobody, so on a public URL it would admit anyone.
Access control
Holding a valid GitHub identity is not enough. ALLOWED_GITHUB_LOGINS lists
the accounts permitted to use the server, and every message is checked against
it. If the list is empty the server refuses to start — an unset allowlist fails
closed rather than exposing health data to any GitHub user.
Tools
Diagnostics — garmin_status
Activities — garmin_get_activities, garmin_get_activities_by_date,
garmin_get_last_activity, garmin_get_activity, garmin_get_activity_splits,
garmin_get_activity_weather, garmin_get_activity_hr_zones
Analysis — garmin_compare_activities, garmin_get_progress_summary
Health — garmin_get_daily_summary, garmin_get_body_battery,
garmin_get_hrv, garmin_get_sleep, garmin_get_heart_rates,
garmin_get_resting_heart_rate, garmin_get_stress, garmin_get_steps,
garmin_get_spo2, garmin_get_respiration
Training — garmin_get_training_readiness, garmin_get_training_status,
garmin_get_max_metrics, garmin_get_hill_score, garmin_get_endurance_score,
garmin_get_race_predictions, garmin_get_training_summary
Profile, goals, gear — garmin_get_user_profile,
garmin_get_personal_records, garmin_get_goals, garmin_get_earned_badges,
garmin_get_badge_challenges, garmin_get_devices, garmin_get_gear,
garmin_get_gear_stats
Weight — garmin_get_weigh_ins, garmin_get_body_composition
Workouts — garmin_get_workouts, garmin_get_workout,
garmin_get_scheduled_workouts
Writes — garmin_upload_workout, garmin_schedule_workout,
garmin_create_manual_activity, garmin_add_weigh_in,
garmin_delete_activity
Every write tool takes confirm and does nothing until it is true. These
change a real training record that feeds load and fitness metrics, so an
agent should show its intent and get agreement first rather than write
speculatively.
Setup
1. GitHub OAuth app
Create one at Settings → Developer settings → OAuth Apps with the callback
URL <BASE_URL>/auth/callback. Copy the client id and secret.
2. Build and sign in to Garmin (once)
cp .env.example .env # then fill it in
docker build -t mcp-garmin .
# Interactive: asks for password and any MFA code.
docker run -it --rm -v /srv/garmin-data:/data mcp-garmin garmin-mcp-auth
Tokens land in /data/garminconnect with mode 0600 and are refreshed
automatically from then on.
3. Run
docker run -d --name mcp-garmin --env-file .env \
-v /srv/garmin-data:/data -p 8000:8000 mcp-garmin
4. Add the connector
- Claude Code —
claude mcp add --transport http garmin <BASE_URL>/mcp - claude.ai / Claude Desktop — Settings → Connectors → Add custom
connector →
<BASE_URL>/mcp
Both then send you through GitHub to sign in.
Configuration
| Variable | Purpose |
|---|---|
BASE_URL |
Public URL of this server, no trailing slash |
GITHUB_CLIENT_ID / GITHUB_CLIENT_SECRET |
GitHub OAuth app credentials |
ALLOWED_GITHUB_LOGINS |
Comma-separated logins permitted to connect |
JWT_SIGNING_KEY |
Stable key for client tokens; unset means clients are signed out on restart |
DATA_DIR |
Where Garmin tokens live (default /data) |
GARMIN_EMAIL / GARMIN_PASSWORD |
Optional, only to pre-fill the interactive login |
HOST / PORT |
Bind address (default 0.0.0.0:8000) |
Notes
Garmin Connect has no public API; garminconnect talks to the endpoints the
mobile and web apps use, and Garmin changes them without notice. The dependency
is pinned exactly for that reason — upgrade deliberately and re-check the tools
afterwards.
Calls are synchronous, so each one runs in a worker thread; a slow Garmin response cannot stall the MCP event loop. Sessions that Garmin expires server-side are re-established once and the call retried.
Licence
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。