Most health AI apps upload your Apple Health data to a server. This is the alternative: a model that runs on the phone, answers real questions about your readings, and shows the samples behind every answer.
The privacy part is not a feature on a list, it is the architecture. The model runs on your iPhone. Your heart rate, HRV and sleep never leave the device. The question, the computation and the answer are all local. There is no upload and no cloud account in the path.
Provenance is the second half. When the app says your HRV averaged 62 ms this week, it shows the metric, the date range, the raw sample count, the computation and a chart. You can tap through to the individual readings (58, 64, 67, 61, 59) and confirm the math yourself. That is the difference between trusting an app and knowing you are right.
If you run your own agents, the same 190 metrics export as clean JSON and a zero-dependency, read-only MCP server lets Claude, Cursor or opencode query the live data on your machine with no network calls at all.
Full write-up with the provenance breakdown: https://dub.sh/dfEIIKS
Top comments (1)
tr.ee/dev-to