Best Fitness MCP Servers in 2026: 11 Options Compared
Eleven ways to give Claude or ChatGPT your training, sleep and recovery data over MCP, compared on sources, write-back, login method, setup and price.
You open Claude and ask whether you are ready for the long run on Sunday. It gives you a careful, sensible answer about sleep, fueling and listening to your body. It has no idea you slept five and a half hours, that your resting heart rate is up 6 BPM this week, or that Sunday is your third hard session in five days.
The fix for that is not a better prompt. It is a fitness MCP server: a connector that lets the model fetch your activities, sleep and recovery data itself. Search for one and you get a wall of GitHub repos, directories and vendor listicles. This guide sorts them.
We make one of the options below, athletedata, and it is listed first because it is ours, not because we pretend to be neutral. Everything about the others comes from their own websites, pricing pages and repositories, checked on 2 November 2026. Prices change, so confirm on the vendor page before you buy.
At a glance
| Server | Type | Sources | Writes back | Price |
|---|---|---|---|---|
| athletedata | Hosted | 39 apps over MCP: Garmin, WHOOP, Oura, COROS, Polar, Suunto, Wahoo, Zwift, Intervals.icu, TrainingPeaks, Hevy, Apple Health, Health Connect and more | Yes: workouts to Garmin, COROS, Apple Watch and Zwift; Intervals.icu, TrainingPeaks and Hevy calendars; food log | $9/month or $69/year, 7-day free trial |
| freddy | Hosted | Garmin, Oura, Polar, Suunto, Wahoo, Withings and others directly, many more via Apple Health and Health Connect | No, read-only | Free (1 source, 7 days), $49/year |
| Wellness Project | Hosted app with MCP | Garmin, COROS, Fitbit, Oura, WHOOP, Withings, Polar, Apple Health, Health Connect and more | Logs workouts and meals | Free (3 chats a day), $9.99/month |
| Fulcra | Hosted, plus open-source local | Apple Health, location and calendar via its iPhone app | Records data and files | Free tier, Pro $14.99/month |
| Strava MCP | Official, single platform | Strava only | No, read-only | Needs a Strava subscription |
| Garmin Personal API | Official key, you build the server | Garmin only | No push to watch | Free, plus your time |
| garmin_mcp (community) | Local, open source | Garmin only | Yes, including workouts | Free |
| intervals.icu community servers | Local, open source | Intervals.icu | Some read and write | Free |
| Health Auto Export | Runs on your iPhone | Apple Health only | No | Premium from $1.99/month |
| Open Wearables | Self-hosted developer platform | Many, with your own developer credentials | No | Core free, Pro from $999/month |
| Terra | Developer API | Hundreds of devices | No | From $399/month billed annually |
What a fitness MCP server actually does
MCP, the Model Context Protocol, is the open standard that lets an AI client call external tools. A fitness MCP server advertises tools like "get activities between these dates" or "get HRV for the last 30 days", and the model decides when to call them. You ask a question, the model fetches what it needs, and it reasons over your real numbers instead of your memory of them. If you want the longer explanation, the guide to connecting your training data to ChatGPT and Claude walks through it end to end.
That sounds like a commodity. It is not, because the hard part is not the protocol. It is everything behind it.
How it logs in. Some servers use each provider's official partner API with OAuth, where you approve access on Garmin's or WHOOP's own consent screen. Others log in as you, with your email and password, and imitate the mobile app. The second kind works until the provider changes its login, and it means a script somewhere holds your password.
Where it runs. Hosted servers give you one URL that works in Claude, ChatGPT and Cursor from any device. Local servers run on your laptop and stop when it sleeps.
Whether it writes. Read-only is fine for analysis. But the moment you ask "build me tomorrow's threshold session", the useful answer ends with the workout on your watch, not in a chat bubble you have to retype.
Whether it computes anything. Raw records are a start. Fitness and fatigue trends, power and pace curves, intensity distribution and load flags are what turn a pile of activities into an answer. Some servers do that math; with others, the model does arithmetic over thousands of rows and you hope it is right.
Those four questions explain most of the differences below.
Hosted servers for your own data
1. athletedata
What it is. A hosted MCP server built for athletes. You connect your apps on the dashboard once, then add one URL to your AI client. It is listed in Claude's connector directory and works in ChatGPT, Codex, Cursor, Lovable and Perplexity.
Sources. 39 apps are queryable over MCP: Garmin, COROS, Polar, Suunto, Wahoo, Zwift, Rouvy, MyWhoosh, TrainerRoad, Hammerhead Karoo, Peloton, Concept2, Intervals.icu and TrainingPeaks on the training side; WHOOP, Oura, Ultrahuman, Withings, Eight Sleep, Google Health, Apple Health and Health Connect for sleep and recovery; Hevy, Strong, Fitbod and Technogym for strength; Cronometer, MyFitnessPal and others for nutrition. Garmin, WHOOP and Oura come through official APIs, not password logins.
Writes back. This is the main difference from every other hosted option. Your AI can push a structured workout to a Garmin, COROS or Apple Watch, or straight to Zwift, with power targets resolved to watts from your FTP. It can also write into your Intervals.icu or TrainingPeaks calendar, create Hevy routines, and log food. The guide to pushing workouts to your Garmin watch shows what that looks like.
Computes. Fitness, fatigue and form, power and pace curves, race-time prediction bands, intensity distribution, load and injury-risk flags, and official race results for HYROX, IRONMAN and the big marathons.
Not so good. No Strava over MCP, because Strava's terms forbid it (see below). Garmin's API only backfills about 30 days at connection, so older Garmin history needs an export upload. And it is built for people who train; if you want location history or a general life log, Fulcra is closer.
Price. MCP tier $9 a month or $69 a year. A full AI coach that messages you first on WhatsApp, Telegram or iMessage and runs an adaptive plan is $39 a month. Both start with a 7-day free trial.
Best for. Athletes with more than one device who want their AI to analyze and also act.
2. freddy
What it is. A hosted, read-only "personal health" MCP server with iPhone and Android companion apps. You get one private URL and it installs from Claude's directory or as a ChatGPT app.
Sources. Direct connections to Garmin, Oura, Polar, Suunto, Ultrahuman, Wahoo, Withings, Concept2, Hevy and Intervals.icu, and a long tail through Apple Health and Health Connect. WHOOP is on a waitlist at the time of writing.
Good. Cheap, fast to set up, and clear about being read-only. The free tier lets you try it with one source.
Not so good. Nothing goes back to your watch or calendar, and there are no computed training metrics beyond what the sources send. Many sources arrive second-hand through Apple Health or Health Connect, which carry less detail than the provider's own API.
Price. Free with one source and seven days of history, Pro at $49 a year, or a $249 lifetime plan.
Best for. Someone who wants raw health data in their AI cheaply and does not need it to do anything.
3. Wellness Project
What it is. A coaching app with several AI coach personas, a dashboard and logging, which also exposes a hosted MCP server for Claude, ChatGPT and other assistants.
Good. Your assistant can log workouts and meals back into the app. Device coverage is broad and includes WHOOP and COROS.
Not so good. The free tier caps you at three coach chats a day, and MCP use counts against the same cap. The MCP is a side door into its own app rather than the main product, so you end up living in two places.
Price. Free, or Pro at $9.99 a month.
Best for. People who want an app-based coach and also like asking Claude now and then.
4. Fulcra
What it is. A "context lake" that puts Apple Health, location and calendar on one timeline, with a hosted MCP endpoint and an open-source local server.
Good. Strong on privacy controls and sharing slices with a doctor or coach. Agents can write data back. The free tier includes MCP.
Not so good. Everything wearable arrives through its iPhone app reading Apple Health, so there are no direct Garmin, WHOOP or COROS connections, and no training analytics.
Price. Free tier, Pro at $14.99 a month or $149.99 a year.
Best for. iPhone users who care about a whole-life log more than training.
Official single-platform options
5. Strava's own MCP connector
Strava launched an official, read-only MCP connector on 1 June 2026, available to Strava subscribers and to Claude first. It covers your activities, streams, power and GPS.
The important context: Strava's API agreement prohibits other apps from passing Strava API data to MCP servers or third-party AI tools. That is why athletedata does not serve Strava over MCP, and why community Strava MCP servers sit outside Strava's terms. If Strava is your only source and you subscribe, use Strava's connector. If you record on a Garmin or COROS, the same activities are available from the watch's own API, along with the sleep and HRV that Strava never had.
6. Garmin Personal API
In October 2026 Garmin opened an official Personal API to individuals. You create a key and pull your own activities, FIT files and health data. It is the right foundation if you want to build your own server: it is official, it does not need your password, and it is free.
What it is not is an MCP server. You still write and host the code that turns the key into tools. The Personal tier is pull-only with no Training API, so it cannot send a workout to your watch, and it reads Garmin and nothing else. The Garmin Personal API guide covers the build and where it stops.
Local and open-source servers
7. garmin_mcp (community)
The most popular Garmin MCP project on GitHub, MIT licensed, with well over a hundred tools covering activities, sleep, HRV, body composition and more. It can also schedule workouts and edit activities.
The catch is the login. You run a one-time setup with your Garmin email and password (and MFA code), and it stores session tokens on your machine that last about six months. That is the same route the mobile app uses, not Garmin's partner API, so it can break when Garmin changes things. If you are comfortable with that and with keeping a Python process alive, it is the most capable free Garmin option. Our Garmin MCP server guide compares it with the hosted route in detail.
8. intervals.icu community servers
Several open-source servers wrap the official Intervals.icu API, which uses a personal API key from your Intervals.icu settings. Some are read-only; others can write structured workouts into your calendar, which Intervals.icu then syncs to Garmin and Zwift. For cyclists already living in Intervals.icu, this is a strong free route. The intervals.icu MCP server guide covers the options.
Community TrainingPeaks servers also exist, but they authenticate with a browser session cookie against unofficial endpoints, which is fragile.
9. Health Auto Export
An iPhone app that runs an MCP server on the phone itself, exposing more than 150 Apple Health metrics. No account, no cloud.
It has real constraints: your computer must be on the same Wi-Fi as the phone, the app has to stay in the foreground while it serves, and it only sees Apple Health. MCP needs the Premium plan, from $1.99 a month or $24.99 lifetime. The Apple Health MCP server guide covers the hosted alternative.
Built for developers and companies
10. Open Wearables
An MIT-licensed, self-hosted platform that normalizes many wearable APIs, with a small MCP server on top. It is aimed at product teams: you run the backend and obtain each provider's developer credentials yourself, which for Garmin or WHOOP means applying to their partner programs. Core is free; Pro starts at $999 a month per deployment.
11. Terra
A developer API covering hundreds of devices, with an MCP interface for apps built on it. It needs a developer account and API keys and starts at $399 a month billed annually. Great if you are building a health product. Overkill for your own data. Sahha sits in the same category with a scores-and-biomarkers API from $299 a month.
Built in, no MCP needed
ChatGPT Health can read Apple Health and medical records, but only in the US at the time of writing, and Garmin, WHOOP or Oura data only arrives if those apps already write it to Apple Health.
Claude on iPhone can read Apple Health in a US-only beta on paid plans, read-only.
Both are convenient if you are in the US and on an iPhone. Neither works in Cursor or Claude Code, neither reaches the provider APIs directly, and neither is available in Europe yet.
Best by sport and goal
The right server depends on what you train for, because each sport leans on different data.
- Running. Pace curves, race predictions and whether your easy runs are actually easy. See Claude as a running coach.
- Cycling. Power curve, FTP, fitness and fatigue, and getting the workout onto your head unit or Zwift. See Claude as a cycling coach.
- HYROX. Running, erg numbers, strength logs and official station splits in one place. See the HYROX AI coach guide.
- Longevity and healthspan. VO2 max, resting heart rate and HRV trends, body composition and bloodwork. See the longevity MCP guide.
- Strength. Hevy's log of every set, which the Hevy MCP server guide covers.
How to choose
Run through these in order. The first one that bites usually decides it.
- Do you live in Strava only? Use Strava's own connector. Nothing else can legitimately serve that data to an AI.
- Do you want to self-host? Use Garmin's Personal API with your own server, the community Garmin server if you accept the password login, or an intervals.icu server.
- Are you building a product? Open Wearables or Terra.
- Do you want your AI to act, not just answer? You need write-back: workouts to your watch, your training calendar, your log. That narrows the hosted field to athletedata for watch delivery, and Wellness Project for logging inside its app.
- Do you only need raw data, cheaply? freddy is the budget pick.
Whatever you choose, test it with a question you already know the answer to. Ask for last week's total running distance, or your average sleep over the past 14 nights, and check it against your watch app. A server that gets the easy question wrong will get the hard ones wrong too.
Putting it together
- List where your data actually lives: your watch, your recovery wearable, your strength log, your training calendar.
- Rule out options that miss your main source, or reach it only second-hand through Apple Health.
- Decide whether you need write-back. If you want workouts on your wrist, most of the list drops away.
- Check the login method. Prefer official OAuth or an official key over your account password.
- Run the known-answer test above in your AI client before you trust it with a training decision.
If you want to try the hosted route, you can connect your apps and add the athletedata MCP server in a few minutes, with a 7-day free trial.
