The loudest request in fitness software right now is an MCP connector

Journal · 2026-09-15 · R. Klug · All notes

Read in: Deutsch · Français · Español · Italiano

Something unusual is happening on the public wish lists of fitness software. On the public feedback board of Bevel, one of the most popular health platforms, the single most-requested feature — 861 votes as of September 2026, far ahead of everything else — is a Claude / MCP connector: a way for people to let their own AI assistant read their training and recovery data. Strava shipped an official MCP connector in June 2026; COROS became the first watch brand with one; ChatGPT added a health surface. The demand is not niche anymore. Athletes have discovered that a season of real data plus a capable assistant beats any canned insights feed — and now they want the pipe.

What people actually do with it

The use cases in those request threads are strikingly consistent: review the week against the plan; ask why Tuesday felt terrible; compare this season's long runs with last season's; build the next training block from what the record actually shows. In other words — the analysis layer athletes always wanted, but shaped by their own questions instead of a vendor's dashboard. We wrote about that working loop here.

The fork in the road: hosted or local

Every connector has to answer one architectural question: where does it run? Most vendors answer with a hosted service — an OAuth login, a server that sits between your data and your assistant, often a subscription on top. It works, but it inherits everything a service inherits: the account, the uptime, the terms. Recent history shows how fragile that chain is — one major platform revoked its public API with weeks of notice, another locked out unofficial clients overnight. A connector built on someone's cloud API lives exactly as long as that API's goodwill. And athletes have noticed the pricing pattern too; the sharpest comment in those threads calls it paying monthly “just to view your own data.”

The local answer inverts it. If the training record already lives as files on your own device, the connector can be a small, read-only server on your own computer — no relay, no login, nothing between the file and the assistant except your decision to connect them. Nobody can sunset it, because there is no service to sunset.

What ampulse ships

ampulse records physiology as files on your iPhone — full Polar Loop nights, raw RR intervals from chest straps, Concept2 and FTMS sessions. The ampulse connector is the local kind: you link a folder in the app (Settings → Connect Your AI), install ampulse.mcpb in Claude Desktop with a double-click, and its fifteen read-only tools serve exactly the records your questions need — session intervals, night data, heart rate, power, GPS. For other assistants, the AI dossier exports the same record as a Markdown file you can hand to anything that accepts files.

Plainly stated, as always: whatever you ask Claude to read is shared with the assistant provider, like anything you put in a conversation — the difference is that this transfer happens by your hand, per question, from files you can open yourself. The connector is free, and it keeps working whether or not any particular service, API program, or company has a good quarter.

Why this belongs together with raw recording

An assistant is only as good as the resolution of what it reads. A connector that serves pre-computed scores lets your assistant paraphrase a dashboard. A connector that serves the actual record — every beat interval, every window with its quality receipt — lets it do analysis: drift across a long session, heart rate at fixed wattage across a season, the nights around a heavy week. That is the quiet reason recording raw and connecting locally are one feature, not two. The 861 votes are asking for the pipe. The pipe is only worth building to a record that has something in it.