health-mcp

A single agent-facing surface over one person's health data, so an AI can cross-reference sources that usually sit in separate apps — training and nutrition today ("did intake track with volume?"), with room to bring in sleep or other metrics later — in service of actually optimizing toward that person's health goals, not just logging numbers.

It's an MCP server, speaking stdio, backed by one SQLite file. Training data is synced in from external sources — currently Hevy for lifting, with Strava for cardio planned next. Food is logged directly through the agent against a small hand-curated Catalog of Products.

Tools

• find_product(query?) — search the Catalog by name/brand; list it all if query is omitted. Always call this before log_food against a Catalog item.

• add_product(...) — add a Product to the Catalog. Macros are per 100g; source is verified (off a package label) or estimated.

• log_food(grams, product_id? | name + macros, at?) — log that a quantity was eaten, either against a Catalog Product or as a one-off (always recorded as estimated).

• delete_food_entry(id) — remove a Food Log Entry.

• sync_workouts(full?) — pull new workouts, exercise templates, and body measurements from Hevy. Delta by default; full=True re-fetches everything and reconciles deletions.

• query(sql) — read-only SQL (SELECT/WITH only) against the whole schema. There are deliberately no narrow read tools (list_workouts, daily_nutrition, etc.) — see ADR-0006.

The domain vocabulary these tools use (Product, Food Log Entry, Macros, Catalog, Verified/Estimated, Day, Workout, Volume) is defined in CONTEXT.md.

Related MCP server: Apple Health AI Bridge MCP Server

Setup

Requires Python 3.13+ and uv.

uv sync

Configure via a .env file in the repo root (or HEALTH_MCP_* env vars):

HEALTH_MCP_DB_PATH=/path/to/health.db   # defaults to ~/health/health.db
HEALTH_MCP_HEVY_API_KEY=...             # required for sync_workouts / `sync`

Migrations run automatically on every CLI invocation.

Running

As an MCP server (what Claude Code / Claude Desktop launch):

uv run health-mcp serve

This repo is already registered as a project-scoped MCP server in .mcp.json, so any Claude Code session opened here loads it automatically.

Other CLI commands:

uv run health-mcp sync           # pull new workouts from Hevy (delta)
uv run health-mcp sync --full    # re-fetch everything, reconcile deletions
uv run health-mcp normalize      # re-parse stored raw workout payloads, no network

Tests

uv run pytest

Roadmap

Next up is a Strava sync for cardio (runs, rides) alongside the existing Hevy sync for lifting, so query can answer questions across both. See "Future ideas" in PLAN.md for this and other planned work.

Design notes

Key decisions and their rationale live in docs/adr/:

• 0001 — day runs 04:00–04:00 Europe/Amsterdam

• 0002 — food log entries store macros, not references

• 0003 — own the data rather than compose external MCP servers

• 0004 — no Albert Heijn integration in v1

• 0005 — stdio over SSH instead of a public HTTPS endpoint

• 0006 — one read-only SQL tool instead of narrow read tools

• 0007 — Claude Code remote access instead of a dedicated VM