Garmin Connect

garmin-connect-api

Read and write Garmin Connect from an MCP client such as Claude, from Python, or over a local HTTP API: activities with per-set strength detail, daily health data, planned workouts, completed strength sessions, and weigh-ins with full body composition.

Unofficial. Not affiliated with or endorsed by Garmin. It uses the same private endpoints as

the Garmin Connect website, through python-garminconnect,

because Garmin offers no public API for individuals. Those endpoints can change without notice,

and using them is not sanctioned by Garmin's terms. It is meant for your own account and your

own data.

Status

v0.3. Reads activities with per-set strength detail, laps and pool-swim lengths, daily health, trends over up to two years and named personal records; writes strength, run, ride and swim workouts, completed strength sessions, set corrections and weigh-ins with body composition. 38 tools, every one checked against a real account. Bug reports and feature requests go to GitHub Issues.

Quick start

You need uv and Python 3.12 or newer.

1. Log in once, at a terminal:

uvx --from garmin-connect-api garmin login

This is the only place a password is ever typed. It asks for your Garmin email, password and any multi-factor code, then stores tokens in ~/.garmin-api/default, readable by your user only. Nothing else ever prompts for a password; without tokens, every tool says to run this.

2. Add the server to your MCP client.

Claude Code:

claude mcp add --scope user --transport stdio garmin -- uvx garmin-connect-api@latest

Claude Desktop, in claude_desktop_config.json (in ~/Library/Application Support/Claude/ on macOS, %APPDATA%\Claude\ on Windows):

{

"mcpServers": { "garmin": { "command": "uvx", "args": ["garmin-connect-api@latest"] } } }

@latest makes uvx check for a new release each time the server starts. Without it, uvx keeps running the first version it downloaded.

3. Ask. "How did I sleep this week?", "Did my pace hold up in yesterday's swim?", "How has my resting heart rate moved since summer?", "Show my last strength session's sets, and change sets 3 to 5 to dumbbell rows", "Schedule a swim for Thursday: 200 easy, then 8 x 100 free on 20 seconds rest, in a 25-yard pool".

What it can do

38 tools, each annotated so your client knows which only read (24), which add something (8) and which delete or overwrite (6), and can ask before running those. Every argument, down to a workout step's pace target, carries a description.

Four tools are written for a model rather than mirrored from an endpoint: garmin_status checks the login without calling Garmin, find_exercise finds Garmin's name for a lift, and recent_strength_sessions and week_summary answer common questions in one call.

Logging a finished workout is the write Garmin's own developer programme has never offered. A session is encoded as a FIT activity file with set messages and uploaded, so it appears in Garmin Connect as a Strength Training activity, exercise by exercise.

Body composition from a smart scale (fat, water, skeletal muscle, bone, BMI, basal metabolic rate, visceral fat) is written as a FIT weight-scale file, the way Garmin's own Index scales do it. A plain weigh-in carries weight alone.

Example output

Answers come back compact, summarised the way a person would read them. These are from made-up data in the shapes Garmin returns.

"Did my pace hold up in yesterday's swim?" calls get_activity_laps:

{

"sport": "lap_swimming", "pool_length": 25.0, "pool_unit": "m", "summary": { "lengths": 80, "distance": 2000.0, "average_pace": "1:52/100m", "first_half_pace": "1:50/100m", "second_half_pace": "1:54/100m", "by_stroke": {"FREESTYLE": 1800.0, "BACKSTROKE": 200.0}, "fastest_segment": {"lengths": "5-8", "pace": "1:44/100m"}, "slowest_segment": {"lengths": "77-80", "pace": "2:01/100m"} } }

"How has my resting heart rate moved since summer?" calls get_trend, which rolls a long range into weeks:

{

"metric": "resting_hr", "unit": "bpm", "aggregation": "week", "summary": { "days_with_data": 88, "average": 54.2, "min": 49.0, "max": 61.0, "first_period_average": 57.3, "last_period_average": 51.6, "change": -5.7 } }

Before you write anything

• Activities you log may be shared automatically. If your Garmin account is connected to Strava or another app, an uploaded session can be posted there too, and deleting it from Garmin does not delete it there.
• Rate limits are taken seriously. Garmin puts Cloudflare and per-account rate limits in front of its login, and a run of failed logins can lock an account out for a day or more. Nothing here retries a login, and a rate limit stops every call path rather than being retried. If you see one, wait.
• Exercises need Garmin's own names, such as BENCH_PRESS / BARBELL_BENCH_PRESS. Every write but the raw workout DTO checks them against the catalogue first and suggests near misses, so a typo fails with a useful message rather than Garmin's bare 400. Common gym names work in search: "rdl", "ohp", "pull ups", "bench presses".
• To correct a logged session, edit its sets. Logging it again with a set changed creates a second activity; logging the identical session again is reported as a duplicate.
• Sessions and body-composition readings need a timezone offset, or a 7am reading would land hours out. They are refused without one.

Command line

garmin doctor                          # setup check, no network

garmin activities --type strength_training --limit 5 garmin sets # per-set detail garmin day 2026-10-18 # one day of health data garmin exercises "lat pulldown" # search the catalogue, offline garmin workout create workout.json # plan a strength workout garmin log-session session.json # log a completed one garmin encode-session session.json # build its FIT file without uploading garmin decode-fit activity.zip # read sets out of a Garmin export garmin serve # the HTTP API on 127.0.0.1:8765

Run them with uvx --from garmin-connect-api garmin ..., or install with uv tool install garmin-connect-api. Example specs: session.json, workout.json.

An encoded session file can also be imported by hand on the Garmin Connect website, if you would rather nothing logged in on your behalf.

Python library

from datetime import datetime, timedelta, timezone

from garmin_api.services import activities, health, uploads from garmin_api.models.workouts import ( StrengthSessionSpec, StrengthSessionExercise, StrengthSessionSet, )

for activity in activities.list_activities(limit=5, activity_type="strength_training"): sets = activities.get_sets(activity.activity_id) print(activity.name, sets.total_volume_kg, "kg")

print(health.day("2026-10-18").steps)

uploads.log_strength_session(StrengthSessionSpec( name="Push A", start_time=datetime(2026, 10, 18, 7, 0, tzinfo=timezone(timedelta(hours=-4))), exercises=[StrengthSessionExercise( category="BENCH_PRESS", exercise_name="BARBELL_BENCH_PRESS", sets=[StrengthSessionSet(reps=10, weight_kg=60.0, rest_s=120)], )], ))

Responses are pydantic models with Garmin's inconsistencies ironed out: kilograms rather than grams, timezone-aware instants, snake_case throughout. Garmin's own payload is still on raw when something is not mapped.

HTTP API

garmin serve runs the same API on 127.0.0.1:8765, with interactive docs at /docs. The MCP tools are generated from these routes, so the two cannot drift apart. Because it can write to your account, the service refuses writes from web pages open in your browser and requests naming any host but localhost, which defeats DNS rebinding. Binding beyond localhost requires GARMIN_API_KEY, and the host name clients will use must be listed in GARMIN_API_HOSTS. garmin mcp --http serves the MCP tools over HTTP behind the same guard.

Configuration

Docker

The Dockerfile builds the MCP server from a checkout. Tokens live in /tokens inside the container; mount an existing directory there and run as your own user, so refreshed tokens stay yours. Logging in works inside the container too:

mkdir -p ~/.garmin-api/default

docker build -t garmin-connect-api . docker run -it --rm --user "$(id -u):$(id -g)" -v ~/.garmin-api/default:/tokens garmin-connect-api garmin login docker run -i --rm --user "$(id -u):$(id -g)" -v ~/.garmin-api/default:/tokens garmin-connect-api

Development

git clone https://github.com/happyc0der/garmin-connect-api

cd garmin-connect-api uv sync uv run pytest

487 offline tests, needing no network and no account, run on Linux, macOS and Windows. They cover the models against real Garmin response shapes, the FIT codec's round trip, every one of the 1,527 catalogue exercises against the FIT profile, the HTTP routes, and every MCP tool's output against its own advertised schema, which strict clients enforce.

Tests against a real account are opt-in, and the ones that create data need a second opt-in:

GARMIN_LIVE=1 uv run pytest tests/live                      # read only

GARMIN_LIVE=1 GARMIN_LIVE_WRITE=1 uv run pytest tests/live # creates, then deletes

Why the unofficial route

Garmin's official Connect Developer Program is for businesses, has no endpoint that uploads a completed activity, and was closed to new applicants in 2026. The aggregators that wrap it cost hundreds of dollars a month and are read-only or plans-only. So this uses the endpoints Garmin's own website uses, gently: no polling, no retries on a rate limit.

Citing

If you use this in research, cite it with the metadata in CITATION.cff; GitHub's "Cite this repository" button formats it.

Licence

MIT. Garmin and Garmin Connect are trademarks of Garmin Ltd., used here only to say what this works with.