Connect Your Training Data to ChatGPT or Claude

You paste a screenshot into ChatGPT, and it guesses

You had a big week. You open ChatGPT, type "was that too much volume?", and paste a screenshot of your Strava calendar. It gives you a thoughtful, confident, completely generic answer. It cannot see your resting heart rate creeping up. It does not know you slept badly on Wednesday. It is reasoning about a stranger who happens to share your screenshot.

This is the wall everyone hits when they try to use a general AI as a training analyst. The model is smart enough to help, but it is blind. It only knows what you can be bothered to type, and you cannot type three months of data.

There is a clean fix now, and it is the thing that has the connector category buzzing: you give the AI direct, live access to your actual training data. Then "how was last week?" is a question it can answer from your real numbers, not a guess about a screenshot.

Why your AI gives generic advice

A large language model is a reasoning engine with no memory of you. Out of the box, ChatGPT and Claude know nothing about your training. Every conversation starts from zero.

People work around this in three painful ways. They paste screenshots, which the model reads imperfectly and cannot follow up on. They copy-paste numbers by hand, which is tedious and error-prone. Or they export a CSV and upload it, which works once and then goes stale the moment you train again.

None of these scale. The data you would need for good coaching - your last twelve weeks across running, lifting, sleep, and recovery - is exactly the data that is too large and too scattered to hand-feed. So you feed a thin slice, and you get advice calibrated to that thin slice.

The deeper problem is that good training decisions are cross-source. Whether today's intervals are a good idea depends on your training load, your sleep, your HRV trend, and what you did yesterday, sitting in four different apps. A human coach synthesizes those. An AI can too, but only if it can see them. Our data-driven athlete guide makes the broader case for why holding all of it in one place changes the quality of every decision.

What MCP actually is

MCP stands for Model Context Protocol. It is an open standard that lets an AI assistant call external tools - little functions it can invoke to fetch information or take action. When a tool is available, the model decides on its own when to use it.

An MCP server for your training data exposes your history as a set of these tools. One tool fetches recent activities. Another pulls sleep and recovery. Another reads your strength sessions, or your body composition, or your power curve. The AI does not need you to paste anything. When you ask a question that needs data, it calls the relevant tool, gets your real numbers back, and reasons over them.

The shift is from describing your training to letting the AI read it. You ask "did my easy runs actually stay easy this month?" and instead of guessing, the model pulls your runs, checks the heart rate against your zones, and tells you which ones drifted. That is a different category of answer.

What you can ask once it is connected

Once your data is live, the useful questions change. A few that only work when the AI can see real numbers:

• "Compare my training load this block to the same point before my last race."
• "My WHOOP recovery has been red three days running. What should I cut this week?"
• "Has my resting heart rate trended up over the last month?"
• "Did my squat volume actually go up, or did I just add accessory work?"
• "I felt flat in today's ride. Was my sleep debt building before it?"

These are follow-up-able. The AI can pull the next layer when you push on its answer, because the data is there, not in a screenshot you would have to re-take. It behaves less like a chatbot and more like an analyst with your file open.

The other thing that changes is preparation work. Before a hard block, you can ask it to summarize where you are starting from. Before a race, you can have it compare your current fitness to the same point in past build-ups. After a season, you can ask what your most productive months looked like and what they had in common. None of these are questions you would bother answering by hand, because the data-gathering alone would take an evening. With the data live, they are a sentence.

Which apps you can connect

One connection covers a wide spread of sources, so your AI sees the whole picture rather than one silo:

• Training and activities: Strava, Garmin, Wahoo, Intervals.icu, TrainingPeaks, Polar, COROS, Suunto, Zwift, TrainerRoad.
• Strength: Hevy.
• Recovery and wearables: WHOOP, Oura, Apple Health, Withings.
• Nutrition: Cronometer, MyFitnessPal.
• Cycle: Clue, Flo.

You connect each source once to AthleteData, the same way you would for any integration. From then on it is queryable through the single MCP connection. If you already use our Strava integration or Garmin AI coach, that data is ready to expose to your AI client immediately.

How to set it up

Setup takes a few minutes and is a one-time job per client.

First, connect your training apps on the AthleteData dashboard. Then open the MCP page, where you will find your personal API key and a short config snippet. The connector works with Claude, ChatGPT, and Cursor.

You add AthleteData's MCP server to your client's configuration and paste in your API key. In Claude Desktop or Cursor that means dropping the snippet into the MCP settings. The client then discovers the available tools automatically, and you are ready to ask questions. You can see the full setup and grab your key on the MCP connector page.

If a step does not work or a provider you want is missing, the team genuinely wants to hear about it - the MCP page lists direct contacts for exactly that.

MCP tier versus the full coach

There are two products here, and which one fits depends on how hands-on you want the coaching to be.

The MCP tier is for people who already live in Claude or ChatGPT and want their training data available there. It is $9/month or $69/year. You drive the conversation; the AI answers from your data. There is no proactive messaging and no automatic plan.

The full coaching plan is $39/month or $299/year. It adds the proactive Telegram coach that messages you when your recovery drops, builds and rebalances your training plan, and pushes workouts to your watch. It uses the same data, but it comes to you instead of waiting for you to ask. If you have ever weighed coaching against doing it yourself, our AI coach versus personal trainer guide is a useful frame.

Plenty of people start on the MCP tier to get their data into their AI, then move up when they want the coaching to be automatic. Both include a 7-day free trial.

Your data, your AI

A fair question with any connector is who sees what. The model here is simple: your AI client connects through your own API key, which only you hold and can revoke from the dashboard whenever you want. The connector reads your data to answer your questions. It does not write anything back to Strava or Garmin or post on your behalf.

You also choose what is connected. If you only want your AI to see running and recovery, connect those and leave the rest off. The scope is yours to set, and the data stays the same data your coach would use - centralized once, then read wherever you want to reason about it. If the sheer volume of scattered metrics is what has held you back, our guide on beating fitness data overwhelm covers how consolidation makes the numbers usable instead of noisy.

A real conversation, end to end

To make this concrete, here is the shape of a session that only works when the data is live.

You start with "give me a read on the last four weeks." The AI calls the activities tool, pulls your sessions, and comes back with your weekly volume, the split between easy and hard work, and a note that your hard days have crept from two to four a week. Already this is grounded in your real log, not a vibe.

You follow up: "is that why I have felt flat?" Now it pulls your recovery data. It finds your HRV has drifted below baseline over the same stretch and your sleep has averaged forty minutes less than the prior month. It connects the two: more intensity, less recovery, declining HRV. That is a diagnosis, and it came from three tools called across two questions.

You push once more: "so what do I cut?" It looks at which sessions are load-bearing for your goal versus which are filler, and recommends dropping one of the midweek hard days back to aerobic work for ten days, then reassessing. Every step of that exchange leaned on data it fetched, not data you typed. That is the difference between an AI that helps and an AI that guesses.

What it can and can't do

It is worth being straight about the boundaries, because a connector is not magic.

It reads, it does not write. The MCP connector pulls your data to reason over it. It does not post activities, edit your Strava, or change settings in your apps. That is deliberate - read-only access is safer and is all the analysis needs.

It is as fresh as your sync. Data lands in AthleteData when your providers push or sync it, which for most sources is minutes to a couple of hours after an activity. It is not a live feed during a workout. For "how did last week go" this is irrelevant; for "what is my heart rate right now" it is the wrong tool.

And it is an analyst, not an autonomous coach. In the MCP tier, you drive. It answers what you ask and follows where you push, but it will not message you on its own when your recovery craters. That proactive layer is the full coaching plan's job. Knowing which of those two you actually want is the main decision here, and it usually comes down to whether you would rather pull insight on demand or have it pushed to you.

Putting it together: a setup playbook

1. Connect your training apps to AthleteData from the dashboard - start with the two or three you actually rely on.
2. Open the MCP connector page and copy your personal API key and the config snippet.
3. Add the AthleteData MCP server to your client (Claude, ChatGPT, or Cursor) and paste in your API key.
4. Let the client discover the tools, then ask a real question, like "how did my training load trend over the last month?"
5. Push on the answer. Ask the follow-up. The AI can pull the next layer because your data is live, not pasted.
6. Connect more sources as you go. Each one you add becomes queryable in the same conversation.
7. If you find yourself wanting the AI to message you first, rather than waiting to be asked, move up to the full coaching plan.

The point is small but it changes everything downstream: an AI that can see your data stops guessing. You can start a 7-day free trial and have your training history live in your own AI the same afternoon.