Connecting an AI to your Garmin training data
An assistant that cannot see your completed sessions can only theorise. Once it reads your activities, it compares planned against actual and adjusts. Here are the ways to give it that access, what each is really worth, and what it exposes about you.
Last updated: 2026-08-24
Why data access changes everything
Without data, an assistant produces a generic plan: it knows neither what you absorbed last week, nor whether the last threshold session cost more than expected. You have to restate everything in every conversation.
With read access, the conversation changes in kind: it notices cardiac drift on a long run, spots that a target pace was never held, and proposes an adjustment rather than a repeat.
The available methods, worst to best
Four approaches circulate. They do not demand the same effort, and above all not the same trust.
- Pasting an activity summary into the conversation. Free, immediate, but to be redone every time and limited to what you retype.
- Exporting a file and uploading it. More complete, but the assistant only sees a snapshot, and the operation repeats.
- A community MCP server reading your account live. Comfortable, but it asks for your Garmin password.
- An OAuth-authorised MCP server, reading only what you explicitly allowed, never holding your credentials.
The shortcut that misleads: asking an assistant to "log into Garmin Connect" does not work. Assistants have no Garmin account. They need an intermediate tool, and that tool is what deserves careful choosing.
The first two methods are fine for a trial. Only the last two support real tracking over time.
What an assistant can actually draw from it
The point is not to narrate the session — your watch already does — but to cross-reference what you do not look at.
- The gap between the prescribed target and the pace actually held, rep by rep.
- Load accumulated over three to four weeks compared with the current week: the ratio that best predicts injury.
- Heart-rate drift at constant pace, a marker of fatigue or heat before the clock moves.
- The balance between disciplines for a triathlete, where a neglected sport rarely shows up in how you feel.
These readings exist in specialist platforms. What the assistant adds is doing them in plain language, on demand, and finishing with a training decision.
What you expose, and how to limit it
Training data is health data: heart rate, weight, sleep, the geography of your outings. It deserves the caution of a medical record.
Three habits are enough. Prefer an authorisation revocable from your Garmin account over a tool holding your password. Check what the tool actually reads rather than what it could read. And know where the data is stored, and for how long.
Frequently asked questions
Can ChatGPT or Claude connect to Garmin Connect by themselves?
No. An assistant has no Garmin account and cannot authenticate on your behalf. It needs an intermediate tool — an MCP server or a connector — that holds the authorisation and exposes the data as actions.
Do I need a particular Garmin subscription?
No. Access to your own account data does not depend on any extra subscription. What matters is the authorisation you grant to the intermediate service.
Does my data end up with the AI provider?
What reaches the assistant is the tool's result, inside the conversation. In other words, whatever you ask it to read passes through. That is a reason to prefer tools returning summaries rather than complete raw exports.
How do I cut access later?
With an OAuth authorisation, from your Garmin account, in two clicks, without going through the service. With a tool holding your password, you have to change that password, which disconnects everything else along the way.
Let your assistant write your sessions
Running Workout connects Claude, ChatGPT or Gemini to your Garmin account through the official API. Nothing to install, and no Garmin password to hand over.
See the connection guide