Garmin Connect+ and Strava AI have the right idea, but the Results
are embarrassing
Connecting your Garmin or Strava data to Claude or ChatGPT gives you much better answers and allows you to use much more powerful models.
Here's how to do it easily.
How to Set It Up
- Go to stats.training and subscribe.
- Log in with Garmin or Strava.
- Copy your personal MCP URL and paste it into Claude or ChatGPT connector settings.
That's it. Now you can call Garmin or Strava tools directly from your favorite AI dashboard.
What Your AI Can Actually See
The connector exposes 25 tools across the same categories you'd normally browse inside Garmin Connect.
Activities
- Splits, pace, heart rate, power, individual lap details
Sleep
- Sleep stages, sleep score, restless moments, SpO₂, overnight respiration
Recovery
- HRV trends, Body Battery, resting heart rate, daily stress
Training Load
- Training status, training readiness, morning readiness, hill score
Performance
- VO₂ max, fitness age, race predictions from 5K through marathon
Weekly Aggregates
- Steps, intensity minutes, stress totals
For Strava users, you also get activity feeds, athlete stats, training zones, and stream data (heart rate, pace, cadence per second).
The Interesting Use Cases
The novelty wears off quickly if you only ask: "Show me my last run."
Where this becomes useful is in the kinds of questions Garmin's UI doesn't answer well.
Compare Long-Run Trends
"Compare my last four long runs: pace, average HR, and HR drift across each one."
Garmin only shows one activity at a time. Claude can line up all four runs and tell you whether cardiac drift is shrinking, a strong sign your aerobic fitness is improving.
Spot Recovery Issues Early
"What's the trend in my resting heart rate over the last 30 days? Anything unusual?"
A creeping resting heart rate often signals under-recovery before you actually feel it.
Explain HRV Changes
"My HRV tanked last night. What did I do yesterday that might explain it?"
Claude can cross-reference sleep, training load, stress, and the previous day's workout. Usually the answer is something simple: poor sleep, a late workout, accumulated fatigue. But it's faster than digging through five separate screens.
Sanity-Check Race Predictions
"Based on my last 8 weeks, what 10K time should I realistically target?"
Garmin gives you a race predictor number. Claude can compare it against your actual workout paces and tell you whether the prediction looks realistic or overly optimistic.
Generate Recovery Weeks
"Build me a recovery-week plan based on this week's training load."
This is where conversational AI beats dashboards. You can refine the plan interactively instead of accepting a static template.
Evaluate Ramp Rate
"Look at the last 4 weeks. Am I ramping volume too aggressively relative to my readiness scores?"
The "10% rule" is just a heuristic. Looking at your actual load and recovery trends is far more useful.
Build Race Plans Dynamically
"I have a half marathon in 6 weeks. Given my current fitness and weekly pattern, write me a build-and-taper plan."
You probably won't follow it perfectly, but it's a much better starting point than a generic training plan from a magazine.
Find Hidden Correlations
"Do I sleep worse after interval sessions? Check the last two months."
This is the kind of correlation question dashboards rarely answer well. Sometimes the fix is as simple as moving hard workouts earlier in the day.
What About Privacy?
Nothing about your training data lives permanently on stats.training's servers. When Claude or ChatGPT needs data:
- It sends a tool request.
- The bridge fetches the data from Garmin or Strava.
- The result gets returned directly.
The conversation itself never passes through the bridge, only the requested data.
When you're done, revoke access from Garmin Connect or strava.com/settings/apps. The connection immediately stops working.
The Real Point
Dashboards tell you things. Chat interfaces let you ask things, including the half-formed questions you'd normally ask a coach after a workout.
Garmin and Strava already hold the data. Claude and ChatGPT are already good at reasoning about it. The missing piece was simply a bridge between them.
The most valuable queries are usually:
- Trend comparisons across weeks
- Recovery analysis
- Planning decisions that depend on recent context
- Correlation questions between sleep, stress, and training
Start there.
Try it: stats.training
Top comments (1)
The framing at the end — "dashboards tell you things, chat interfaces let you ask things" — is the whole game with MCP connectors. Most Garmin/Strava analysis tools are stuck showing one activity at a time because that's what fits on a card. The interesting questions almost always span a window of activities and pull in adjacent data (sleep, HRV, stress), which is exactly what an LLM is suited to.
A couple of practitioner notes from using Claude with MCP connectors more generally:
Opus is genuinely good at the cross-tool reasoning step here — the moment where it pulls activities, then sleep, then HRV, and synthesises a single answer rather than dumping three tables at you. That's the part you'd find tedious to do by hand in the Garmin UI, and it's the part the model adds the most value to.
A prompt that consistently raises answer quality on health/training data: ask the model to state its working before the conclusion ("first list the data you pulled, then describe the trend, then give the recommendation"). It anchors the answer in the actual numbers rather than letting the model jump to a confident-sounding generality. Especially important for anything that influences training decisions.
The privacy model you describe — conversation stays with the AI, the bridge only proxies tool calls — is the right default for this kind of personal data, and worth flagging clearly to anyone considering it. "My data never sits on your server" is a meaningfully different posture from how most fitness aggregators have historically worked.
Nice writeup; the HRV/recovery-week use cases are exactly the kind of half-formed questions a dashboard can't answer.