On-device health AI promises privacy, then hands you a chat box. The paradox is that the architecture keeping your data out of the cloud also keeps it out of every other agent.
A July 2026 sweep of the on-device health market found nine apps claiming local processing. All nine had zero MCP support. None shipped a bridge, an export pipeline, or any way for a second tool to read the data.
Here is the part that matters. The privacy win and the portability loss are the same decision. If the only thing that can read your metrics is one app's model, you did not get privacy, you got a smaller silo.
An MCP bridge is the fix. It is a small server that runs on your own machine, reads the export folder directly, and exposes the metrics as live tools any MCP-aware agent can call. Nothing is uploaded. The model stays on the phone. The 14 read-only tools are the same ones Claude, Cursor and opencode already know how to call.
MetricBridge ships both halves: the on-device app that writes the export, and the open-source server that hands it to the rest of your stack. 190 metrics, provenance on every answer.
Top comments (0)