Pass-Through vs Answer-On-Device: Who Computes Your Health Answer
Two designs showed up in health AI in 2026, and they disagree on one thing: where the computation happens.
A pass-through tool moves your data and stops. Your agent gets the samples and does the reading, so the model's arithmetic becomes the product. You audit it by asking again and hoping for the same number.
An answer-on-device app computes the number on the phone first. A model only words the result, and every answer carries a provenance card naming the samples, the date range and the computation behind it.
Why it matters
It is invisible in a demo. It shows up on the third question, when a trend flips sign between two identical prompts. A pass-through pipe has no artefact to inspect, so you cannot separate a data problem from a reasoning problem after the fact.
The design is not either/or
Answer-on-device is the pipe plus a checked answer. MetricBridge exports all 194 metrics as local JSON and CSV, and the open-source MCP server exposes them to any MCP-aware agent as 18 tools (16 read-only, plus add_event and remove_event, which write only an events file beside the export) and 22 prompts. Nothing writes to Apple Health.
So you get a computed answer you can audit in the app, and raw data your own agent can query over MCP.
How to choose
Pick pass-through if you trust your agent and only want a pipe. Pick answer-on-device if you want the number to be right before the model touches it. Pick both if you can, because the interesting questions are the ones where a checked answer and an open agent disagree.
Full post, with the provenance card example and the method behind every computed number: https://www.healthexport.dev/blog/pass-through-vs-answer-on-device?utm_source=devto&utm_campaign=exp-hea525-20261101
Top comments (0)