All Cases

Turning Garmin data into answers, in any AI assistant.

Garmin collects a huge amount of health and training data, then shows it back as charts. I wanted answers instead. So I built an MCP server that lets any AI assistant read my Garmin data and answer questions about it, either on my own computer or from my phone straight after a run.

-99%
Response size
28
Read-only tools
1 line
Local install
Free
To run, local or cloud
Projecting a 10k time in the Claude app
The Challenge

Years of sleep, heart rate and training data, and no good way to ask it a question.

Garmin Connect is built for looking at charts, not answering questions like "Am I ready to train hard today?" or "How much of my running is easy vs hard?".

Most Garmin MCP servers predate Garmin's official Personal API. They log in with an email and password, call private web endpoints, break whenever Garmin changes its login or MFA, and run only on the machine they are installed on, so they never reach web or phone apps. If that password leaks, so does the whole Garmin account.

A basic one-tool-per-endpoint wrapper is no better: Garmin returns 100+ fields per activity and a heart-rate reading every two minutes, so five months of runs comes back as around 190,000 characters, which is slow and expensive for a model to read.

Client
Open Source
Category
Custom AI Tooling
Role
Builder
Year
2026
The Approach

Use the official API, do the heavy lifting in the server, and let people choose where it runs.

The server calls Garmin's official Personal API with a revocable, expiring key, never a password, and every tool is a fixed, read-only request.

Rather than mirroring endpoints, it answers common questions in one call: readiness, wellness trends, training load and personal bests are fetched, joined and aggregated before the model sees them. It splits long date ranges around Garmin's 28 and 31-day limits and fetches them in parallel, resolves periods like "last month" in the user's own time zone, and returns partial results when one source fails. For a single run it reads the activity file from the watch and returns per-km splits, laps and heart-rate drift as a few kilobytes of tables rather than thousands of raw samples.

It also computes what Garmin's API leaves out. There are no personal records in the Personal API, so the server works them out from the user's own activity files: fastest 1 km to marathon over any period. Race estimates use a real recent race time when there is one, and Riegel's formula from the nearest distance otherwise. Training load is tracked week by week, with the change on the previous week and an acute:chronic ratio, and flagged when it ramps up too fast.

A recovery tool checks whether sleep, HRV or Body Battery relate to running performance. A naive comparison said the best-sleep runs were 6% "less efficient", because those were the easy runs. So it compares runs at the same heart rate, and says plainly when there isn't enough data to tell. Catching that kind of confound is the judgement I bring to work like this.

The same 28 tools and 5 prompts then run in two ways. Locally, the user's AI app starts the server on their computer with nothing listening on the network, installing it is one line, npx -y garmin-mcp, and a cache carries over between sessions while the Garmin key and token never touch the disk. In the cloud, it runs on the user's own Cloudflare Worker behind GitHub sign-in and OAuth 2.1, locked to a single GitHub account, so it works in Claude and ChatGPT on the web and on a phone.

Expertise
AI Strategy · Product Design · Development
Tooling
TypeScript · Node.js · Cloudflare Workers · MCP · OAuth 2.1 · GitHub · Garmin Personal API · Claude · ChatGPT
Official API
Server-side Analysis
Local or Cloud
Any AI Client
The Impact

Questions about health and training answered in one call, on any device, at no extra cost.

Five months of runs now comes back as about 1,000 characters instead of 190,000, so answers are fast and cheap in tokens. In the recording below, Claude on a phone projects a 10k race time from recent runs, all from one plain-English question.

The 28 read-only tools cover every endpoint in Garmin's Personal API, including the activity file, and ready-made prompts for readiness, sleep, training reviews, weekly check-ins and session comparisons make it useful from the first question.

People who only use Claude Desktop, Claude Code or Cursor can be running locally with one line, and those who want it on their phone or in ChatGPT can set up the cloud version with one command, npm run setup, with nothing to migrate between the two. An outside user running the cloud version reported back simply: "Works excellent!" Both use the AI subscription the user already has and cost nothing to run, and Garmin data goes straight from Garmin to the user's own computer or Cloudflare account, with no third-party service storing it.

Claude tells users when a new version is out, and updating is also one command, npm run update. The check downloads a small version file from GitHub once a day, sends nothing about the user or their data, and can be turned off. Anonymous download counts of those files are the only usage measure: there's no tracking.

The project is open source under the MIT licence, with every deploy gated on type checks, unit tests and an end-to-end test of the full OAuth and MCP flow. Releases are automated: one GitHub release publishes to npm with provenance, to the official MCP Registry and as a signed Claude Desktop extension.

190k → 1k
Characters for five months of runs
28
Tools, every Personal API endpoint
1 line
To install locally
0
Passwords stored
Sleep
HRV
Activities
MCP server
Am I ready to train hard today?
ClaudeChatGPT
-99% response sizeFree hosting
Projecting a 10k time in the Claude app

The same pattern works for any business with valuable data stuck behind an API or a dashboard: make it answerable in the AI tools your team or customers already use, securely and without a rebuild. If that's you, let's talk.

Book a Call

Let's start building, together.

A 30-minute chat, no pitch deck. Directly with Colin, no middlemen. Honest advice on fit, scope and cost.

1
Date & Time
2
Your Details
November 2026
Su
Mo
Tu
We
Th
Fr
Sa

Select a date