Garmin Training MCP

Читайте данные тренировок Garmin и записывайте структурированные тренировки обратно на часы. Пробежки, отрезки, сон и пульс на входе; беговые и силовые сессии планируются на устройство. Работает локально через stdio или размещается с OAuth 2.1.

GitHubПопробовать этот MCPСпонсировано

Документация

garmin-mcp

Ask Claude about your Garmin data, then have it write the session onto your watch.

A local MCP server that connects Claude Desktop to your own Garmin Connect account. It reads your runs, splits, heart-rate zones and daily health metrics — and, unlike the read-only Garmin integrations out there, it can build a structured workout and schedule it, so the answer to "what should I run on Thursday?" ends up on your wrist instead of in a chat log.

Everything runs as a local subprocess on your machine. No hosting, no server holding your credentials, no network exposure.

You:    My last three runs are all at the same effort. Give me something harder
        for Thursday, based on what my recent paces actually support.

Claude: [reads your activities and splits, then proposes]

        Thursday Threshold (running, about 52m 55s)
          warmup: 15m
        5 x
            interval: 1.00 km @ 4:00/km-4:10/km
            recovery: 1m 30s
          cooldown: 10m

        Create this and put it on Thursday?

Install

macOS 10.15 (Catalina) or newer, with Claude Desktop already installed:

curl -fsSL https://raw.githubusercontent.com/Bartolome69/garmin-mcp/main/scripts/bootstrap.sh | bash

That fetches the code to ~/garmin-mcp, installs a modern Python via uv, signs you in to Garmin, and registers the server with Claude Desktop. It's the only command most people need. Read it first if you'd rather — scripts/bootstrap.sh is short.

Then quit Claude Desktop completely (⌘Q) and reopen it.

Manual install, or on Linux

git clone https://github.com/Bartolome69/garmin-mcp.git
cd garmin-mcp
./scripts/setup.sh          # venv + dependencies
./scripts/login.sh          # sign in to Garmin once, caches the session

Then register it with your MCP client. For Claude Desktop on macOS, ./scripts/install-claude-desktop.sh does it (with the app quit). For anything else, copy .mcp.json.example, fill in the absolute paths, and point your client at python -m garmin_mcp over stdio.

Tools

Tool	What it returns
`get_activities(limit, start_date, end_date)`	Runs, rides and workouts: distance, duration, pace per km and mile, average and max HR, HR zones, cadence, training effect, plus running dynamics and power when the watch records them
`get_activity_details(activity_id)`	One activity in detail: per-split distance, pace, HR, cadence, running dynamics and power, plus full heart-rate time-in-zone
`get_daily_summary(date)`	Steps, distance, calories, resting/min/max HR, body battery, stress, intensity minutes
`get_sleep_data(date)`	Sleep stages with durations and percentages, sleep score, overnight HRV, resting HR
`list_workouts(limit)`	Structured workouts saved in the account
`create_workout(name, steps, sport, description)`	Builds a structured workout and adds it to Garmin Connect
`schedule_workout(workout_id, date)`	Puts a workout on a date, which is what syncs it to the watch
`unschedule_workout(date, schedule_id)`	Takes a workout off a day. Reversible — the workout itself is kept
`delete_workout(workout_id, confirm)`	Deletes a workout. The first call only reads its name back; removing it needs a second call quoting that name
`get_plan_chart(weeks_back, weeks_forward)`	Draws the plan as an image for the conversation: each day's sessions beside what was planned
`get_progress(weeks)`	Which planned sessions actually got done, week by week. Counts a session if it happened within a day either side of its scheduled day
`get_profile()`	VO2 max and personal records — 5k, 10k, half, marathon and the rest
`get_connection_status()`	Whether the server is signed in, which account (masked), and the state of the token cache

Dates accept YYYY-MM-DD, today, yesterday, tomorrow, or a signed offset like -7 or +3.

Writing workouts

create_workout takes an ordered list of steps. Each has a type (warmup, interval, recovery, rest, cooldown or repeat), exactly one of duration_seconds or distance_meters, and an optional target — either pace (minutes per km, as "4:05" or a range ["4:00","4:10"]) or hr ([150, 165]).

15 minute warmup, 5×1km at 4:05 with 90 second recoveries, 10 minute cooldown:

[{"type": "warmup", "duration_seconds": 900},
 {"type": "repeat", "times": 5, "steps": [
     {"type": "interval", "distance_meters": 1000, "pace": "4:05"},
     {"type": "recovery", "duration_seconds": 90}]},
 {"type": "cooldown", "duration_seconds": 600}]

A single pace is widened by 5 s/km either side, because Garmin alerts on a range and an exact target beeps constantly. Repeat groups don't nest. Creating a workout only saves it — schedule it on a date for it to reach the watch.

What it can and can't do to your account

Reading is unrestricted. Writing is confined to the workout library: create a workout, schedule it, unschedule it, delete it. Nothing reaches your training history — no tool deletes or edits a recorded activity, so a run you did cannot be lost here however wrong a tool call goes.

delete_workout is gated rather than trusted to a docstring. The first call never deletes: it reads the workout back and returns its name, and only a second call passing that name as confirm removes it. A model cannot destroy a workout it has not first named.

Your password is read from the environment, sent straight to Garmin, and never written to disk. Only the session token Garmin issues is cached, at ~/.garmin-mcp/tokens.json, written 0600 inside a 0700 directory. No tool returns the password or the token — get_connection_status reports a masked address and the cache's age and permissions, nothing more. Logs go to stderr, so they never corrupt the MCP stream on stdout.

After the first sign-in the server runs off the cached token. Set GARMIN_EMAIL and GARMIN_PASSWORD in the server's environment if you want it to re-authenticate unattended when that token eventually expires; leave GARMIN_PASSWORD out and you'll re-run scripts/login.sh instead.

If it doesn't work

Start here — it walks from the interpreter to a live Garmin call and names the first broken link:

./scripts/doctor.sh

"Garmin is rate-limiting logins from this IP address (429)" — the most common failure, and it isn't your password: Garmin blocks by network address before it checks credentials. Office wifi, university networks and VPNs get hit hardest. Sign in once over a phone hotspot; afterwards the cached session is used instead.

"Garmin is asking for a multi-factor code" — the server can't prompt over stdio, so run ./scripts/login.sh in a terminal once. It handles the code and caches the session.

Claude can't see the tools — Claude Desktop loads its config at launch and writes its own copy back when it closes, so a change made while it's running disappears. Quit it fully, run ./scripts/install-claude-desktop.sh, reopen.

Install fails mentioning Rust, OpenSSL or a compiler — macOS is too old. The hard floor is 10.15 (Catalina), set by the Python interpreter itself: uv's Intel build is compiled with minos 10.15 and will not launch below it, so no amount of package pinning helps. The installers pass --only-binary :all: so this fails fast rather than becoming a doomed source build, and retry against constraints-legacy.txt in case a package has simply dropped a wheel.

No sleep data — the watch wasn't worn overnight, or hasn't synced. Sleep, HRV and overnight body battery only exist if you sleep in it.

Training skills

skills/ holds training methodology that builds on the tools — how to read the data and what to prescribe from it. The first is a half marathon skill in Pete Pfitzinger's style.

Development

.venv/bin/python tests/smoke_test.py

Drives the server over real stdio like an MCP client would, against a stubbed Garmin account — no network, no credentials. Covers every tool's response shape, workout construction, bad input, and the no-credentials startup path.

.venv/bin/python -m garmin_mcp.check

The same code path against your real account, printing what comes back. Useful for confirming a setup end to end.

Other MCP clients, and ChatGPT

Nothing here is Claude-specific: it speaks MCP over stdio, so any client that launches a local server will run it — Claude Code, Cursor, VS Code, Zed, Windsurf. Copy .mcp.json.example, fill in absolute paths, point the client at python -m garmin_mcp.

ChatGPT can't run this. Its connectors take a public HTTPS URL over SSE, because ChatGPT executes on OpenAI's servers and cannot start a process on your machine — there's no local-server option to enable. Serving it to ChatGPT would mean hosting it publicly and holding users' Garmin credentials, which is exactly what this project avoids.

Caveats

Not affiliated with Garmin. It uses the same private API the Garmin Connect website does, via garminconnect, because Garmin publishes no consumer OAuth API. That API can change without notice and take this with it.

MIT licensed. Built with Claude Code.