Which wearable and fitness platform APIs can an independent developer actually get access to — and what do their terms say about handing that data to a language model?
This is field research, not a link collection. It was gathered while building a multi-provider health data integration in 2026, at the point where the obvious question became unavoidable: can I go direct to the providers instead of paying an aggregator?
The short answer was no, for most of them. The long answer is below.
Status: Research as of July 2026. Access programs and terms of service change, some of them specifically to address AI use. Verify against each provider's current documentation before relying on any of this. See DISCLAIMER.md.
| Provider | Self-service access | Verdict |
|---|---|---|
| Whoop | Yes — capped at 10 users until app review | Open now — the one provider you can start building against today |
| Terra (aggregator) | Yes — dashboard signup, paid | The pragmatic door — and for most providers below, the only one |
| Polar | AccessLink developer program | Buildable — the one native integration that was actually shipped |
| Suunto | No — partner approval, companies only | Gated — apply, and bring your own FIT parser |
| Garmin | No — program closed to new applicants | Blocked — watch for it reopening |
| Coros | No — invite-only partner program | Blocked |
| TrainingPeaks | No — waitlist, qualitative selection | Blocked |
| Oura | No | Legally blocked — terms restrict LLM use to Oura's own MCP server |
| Strava | No | Legally blocked — terms prohibit ingestion into a context window |
The full comparison, with OAuth details, webhook behaviour, raw file availability and the AI/LLM clauses, is in patterns/provider-matrix.md.
The lead time is the approval process, not the code. Writing an adapter against a documented API is a few days. Getting credentials can take weeks, or never happen at all. Sequence the work accordingly — apply first, build while you wait.
"AI clause" is now a real column in this table. Several providers amended their terms during 2025–2026 specifically to address language models. Oura restricts LLM use to its own MCP server. Strava prohibits ingestion into a context window outright. Garmin requires transparency and explicit consent rather than prohibiting it. This determines whether your integration can legally exist, and it is not something you can retrofit.
Aggregators are priced like the only door, because they often are. That is a fair description of the value, not a complaint — but it means the build-vs-buy calculation is not the one you would expect, and it does not improve at small scale.
Providers — one document each, covering access, auth model, data model, webhooks, limits, and the terms that matter:
Garmin · Coros · Oura · Polar · Strava · Suunto · Terra · TrainingPeaks · Whoop
Patterns — what generalises across providers:
- provider-matrix.md — the full comparison table
- adapter-interface.md — the abstraction that belongs in place before provider number one, not after number three
- mcp-for-health-data.md — designing MCP tools for a model that has to pay for every answer out of its context window
Anything not verified against a primary source is marked UNCONFIRMED rather than
smoothed into a confident claim. Those markers are the point. A guide that reads
authoritatively about terms nobody actually re-read is worse than no guide.
Where a document says a program is closed or a clause exists, it is because that was observed at the time stated — not because it was inferred from a forum post.
polar-sync— the Polar AccessLink pipeline as a working library: OAuth, FIT parsing, normalization, SQLite/Postgres storagepace-mcp— the archived MCP server this research came out of
Corrections welcome, especially "this changed, here is the current source". See CONTRIBUTING.md.
MIT. See LICENSE.