Workout Detection API: Get Recorded Workout Sessions

A "workout detection API" gets you a user's recorded workout sessions — a run, a ride, a strength set — with their start time, end time, and activity type. You read them two ways: from an on-device store (Apple HealthKit's HKWorkout on iOS, Android Health Connect's ExerciseSessionRecord on Android) or from a cloud activity API (Strava, Garmin, Fitbit) after the user's device syncs. The important split is auto-detected vs. manually started: some vendors silently recognize activity (Fitbit SmartTrack), while others generally expect the user to start and stop the workout. Best pick: for automatic capture, source from a wearable/vendor that does auto-detection; to just read whatever the user already logged, read the on-device store for the broadest coverage. (This page is about recorded sessions, not AI camera-based rep counting — for that, see AI workout tracking APIs.)

Where you can get workout sessions

Which activity types a source recognizes, and whether it auto-starts, are vendor-specific and change often, so treat this table as a starting point and confirm against each provider's live docs (as of 2026, verify).

Auto-detect vs. manual start

This is the differentiator to lead with, because it decides whether your app can rely on workouts appearing on their own:

• Fitbit SmartTrack auto-recognizes select activities after roughly 15 minutes of movement (a configurable threshold, commonly cited around 10–90 minutes) and logs them with no manual start. Which activity types it recognizes changes — verify the current Fitbit help list rather than hardcoding it.
• Apple Watch generally relies on the user starting and stopping a workout (with reminder prompts to start or end), rather than fully silent logging. Verify current watchOS behavior before claiming auto-detection.
• Garmin and Strava records reflect activities the device or app captured; auto-start behavior varies by device and setting — verify per device.

A second thing to plan around: activity-type taxonomies differ across sources. HealthKit's workout activity types, Health Connect's exercise session types, and Strava's and Garmin's activity types are not a 1:1 mapping, so if you merge sources you must map their type enums explicitly rather than assume equivalence. Confirm each taxonomy in the provider's data dictionary.

Measured or estimated?

A recorded workout mixes measured and modeled parts, and it is worth being explicit about which is which:

• Duration is measured — the session's start and end boundaries reflect real elapsed time.
• Activity type and session boundaries are algorithmically inferred, especially for auto-detected workouts. Do not present a detected activity type as certain; SmartTrack classifying a session as "walk" vs. "run" is a model's guess.
• Any derived calories or effort are estimates, modeled from heart rate, motion, and the user's profile — see how calorie tracking burns are modeled. These disagree across devices, so never publish an accuracy figure.

Note the distinction this page is built on: a recorded workout session (what HealthKit, Health Connect, Strava, and Garmin give you) is not the same as AI camera-based rep or exercise detection, which classifies movement from a video/pose feed in real time. That is a separate capability covered under AI workout tracking APIs — don't conflate the two.

Which source should you pick?

• You need workouts captured automatically: source from a wearable/vendor that does auto-detection (for example Fitbit SmartTrack). Verify each vendor's current capabilities, because auto-detection support and the recognized-activity list change.
• You just need to read what the user already logged: read the on-device platform store — HKWorkout on iOS, ExerciseSessionRecord on Android — for the broadest coverage and no post-sync latency, since it inherits whatever apps wrote into the OS.
• You want route and per-workout detail across many athletes: Garmin's Activity API delivers standard .FIT/GPX/.TCX files; use Strava only after confirming its display and usage restrictions fit your product.
• You want many brands behind one schema: an aggregator (Terra, Junction, Rook) normalizes workouts across sources so you avoid N integrations.

Before you ship

Auto-detection availability, the list of recognized activity types, activity-type enums, scopes, and field names are all volatile — re-verify each against the vendor's current docs as of 2026. Map activity-type taxonomies explicitly when merging sources, treat a detected activity type as a model's inference rather than fact, and remember that camera-based rep detection is a different capability entirely.