wearable-mcp — Model Context Protocol server for wearable readings

A Model Context Protocol server that exposes user-consented wearable readings (heart rate, skin temperature, HRV SDNN, 24-hour step count, body temperature) to LLM clients (Claude Desktop, Claude Code, the Sentinel Triage Agent, etc.).

Built for the EpiHack Arizona 2026 Heat focus group as part of Phase 4 of the roadmap.

Privacy stance (read this first)

This MCP server never talks to Apple HealthKit or Android Health Connect directly. Those data sources require on-device user consent and stay on the user’s device. Instead, the server is a consumer of the on-device store that app/shared/wearable.js populates after the user grants permission through the webkit.messageHandlers.health bridge (iOS) or navigator.healthConnect (Android). The MCP-side picture is therefore read-only and downstream of consent.

Until that on-device store exists (it’s a Phase-4 ship), the server runs mock-by-default: every tool returns canned readings drawn from wearable_mcp.mock_data so the demo + tests work without a real device. Set WEARABLE_BACKEND_URL to point at a real consented-store proxy when one is available.

What it does

MCP tool Returns
wearable_recent_readings List of normalised readings for one LOINC code since a given ISO timestamp.
wearable_summary_24h min / max / mean / count for the last 24 h.
wearable_alert_check Which of a small rules-spec fire on the current data (tachycardia, etc.).
wearable_supported_metrics The LOINC codes this build supports, plus human-readable name + unit.

Plus two MCP resources:

Supported metrics (LOINC)

LOINC Name Unit
8867-4 Heart rate bpm
8310-5 Body temperature degC
8328-7 Skin temperature degC
80404-7 Heart rate variability SDNN ms
41950-7 Steps in 24h steps

Why this matters for EpiHack

When the Triage Agent is reasoning about a heat self-report it can ask, in plain language, “is this user currently meeting a tachycardia threshold?” and get a structured deterministic answer from wearable_alert_check, without ever touching the user’s raw stream or leaving the consent boundary the wearable shim already established.

Run

uv run --from . wearable-mcp