apple-health-fly-mcp
MCP server that enables LLMs to query Apple Health data such as steps, heart rate, sleep, and workouts via natural language, with secure cloud access through OAuth.
README
Apple Health MCP Server
An MCP server that exposes Apple Health data as tools queryable by LLMs (Claude, Copilot, etc.).
Architecture
exportación.xml (Apple Health)
│
▼
preprocess.py ──► data/*.parquet (run locally on Mac)
│
▼ (sync_to_fly.sh, over plain HTTPS)
Fly.io: fjcabello-apple-health-mcp
├─ wrapper.py (FastMCP app + api_key ASGI gate, port 8080)
└─ /data volume (persistent, 1GB) ─ loaded by server.py
│
▼
apple-health-fly-mcp-worker (Cloudflare Worker, OAuth gateway)
│
https://apple-health-fly-mcp-worker.fjcabello.workers.dev/mcp
The server loads Parquet files at startup (~1-2 s) and caches them in memory. If they are not found, it falls back to parsing the XML directly. server.py is never modified for deployment concerns — wrapper.py wraps it with the api_key auth gate and the admin sync endpoints (see "Deploy to Fly.io" below).
Requirements
- Python 3.11+
pyarrow(for Parquet support)
pip install -r requirements.txt
pip install pyarrow
requirements.txt
mcp[cli]>=1.0.0
lxml>=5.0.0
pandas>=2.0.0
pyarrow>=15.0.0
uvicorn>=0.30.0
Data update workflow
1. Export from iPhone
Health → profile → Export All Health Data → produces a ZIP containing exportación.xml. Unzip it so the file lands at ../apple_health_export/exportación.xml (relative to this repo), or pass a custom path.
2. Sync to Fly.io
cd ~/Personal/apple_health/apple_health_mcp
./sync_to_fly.sh
This single script:
- Runs
preprocess.py(XML →data/*.parquet, ~20-25 files, one per metric type) - Uploads every Parquet file to the Fly volume via
PUT /admin/upload/<filename>?api_key=...(plain HTTPS — SSH/SFTP tunnels to Fly are blocked on this network) - Calls
POST /admin/reload?api_key=...to clear the in-memory cache, so the next tool call picks up the fresh data — no machine restart needed
Both /admin/* routes require the API_KEY Fly secret as a query param (see wrapper.py). The key is read from .fly_secret_local (gitignored) or the environment.
3. Automated sync via Health Auto Export
Instead of manually exporting the Apple Health ZIP, the Health Auto Export
iOS app can push data automatically via a webhook, which upserts directly into
the Parquet files on the Fly volume (no XML/preprocess step needed). This is
defined in server.py (ingest_app) and wired into wrapper.py's router.
Endpoint (calls Fly directly, not through the Cloudflare Worker):
POST https://fjcabello-apple-health-mcp.fly.dev/ingest
GET https://fjcabello-apple-health-mcp.fly.dev/ingest/inspect (last payload received, for debugging)
Authentication: header x-api-key: <secret> or query param ?api_key=<secret>, checked against the INTERNAL_SECRET Fly secret (kept equal to API_KEY for simplicity — set both with the same value). This is intentionally separate from the Cloudflare Worker's OAuth, since the iOS app can only set a header, not go through the OAuth flow.
Configure one automation per data type in Health Auto Export (the app only allows one data type per automation): Health Metrics, Workouts, etc. — pick the metrics listed in config.py → HK_TYPE_MAP. Format JSON, export version v2, incremental date range, header x-api-key set to the shared secret.
Each /ingest call upserts only the metrics/workouts present in that payload (dedup by startDate, or startDate + activityType for workouts) and clears the in-memory cache so the next MCP tool call reloads fresh data — no restart needed.
Deploy to Fly.io
The server runs as a Docker container on Fly.io (app fjcabello-apple-health-mcp, region ams), fronted by wrapper.py (an ASGI router that gates /mcp and /admin/* behind API_KEY, and forwards /ingest* to server.py's own-auth ingest_app — mirrors the proxy.cjs pattern in garmin-connect-mcp). Parquet files live on a persistent Fly volume mounted at /data, not baked into the image.
First-time setup
fly auth login
fly apps create fjcabello-apple-health-mcp
fly volumes create apple_health_data --region ams --size 1 --app fjcabello-apple-health-mcp
fly secrets set API_KEY=$(openssl rand -hex 32) --app fjcabello-apple-health-mcp
fly secrets set INTERNAL_SECRET=<same value as API_KEY> --app fjcabello-apple-health-mcp
Continuous deployment
.github/workflows/deploy.yml auto-deploys to Fly.io on every push to main, using a FLY_API_TOKEN repo secret (fly tokens create deploy --app fjcabello-apple-health-mcp).
Why CI instead of
fly deploylocally: on this network, Fly's remote "depot" builder and SSH/SFTP tunnels hang indefinitely / fail the WebSocket handshake (corporate SSL/proxy inspection). Deploying from a GitHub-hosted runner avoids this entirely.
Seeding / updating data
See "Data update workflow" above — use ./sync_to_fly.sh, not SSH.
Running locally (development)
python server.py
# Listens on http://0.0.0.0:8001/mcp
Optional environment variables:
| Variable | Default | Description |
|---|---|---|
APPLE_HEALTH_DATA_DIR |
./data |
Directory containing .parquet files |
APPLE_HEALTH_EXPORT |
../apple_health_export/exportación.xml |
XML fallback path |
Available MCP tools
| Tool | Description |
|---|---|
health_summary |
Overview of all available data types, record counts and date ranges |
get_steps |
Daily step counts |
get_heart_rate |
Heart rate per day (mean / min / max) |
get_resting_heart_rate |
Daily resting heart rate |
get_sleep |
Sleep analysis by stage (Core, Deep, REM, Awake) |
get_workouts |
Workout sessions, filterable by type and date |
get_body_metrics |
Weight (kg), BMI, body fat %, lean body mass |
get_activity_energy |
Active/basal energy burned, distance, flights climbed |
get_nutrition |
Nutritional intake (calories, protein, carbs, fat) |
query_health_data |
Generic query for any available metric |
All tools accept optional start_date and end_date parameters in YYYY-MM-DD format.
Available metrics for query_health_data
steps, heart_rate, resting_hr, active_energy, basal_energy, distance_walk, distance_cycling, flights_climbed, sleep, body_mass, bmi, body_fat, lean_body_mass, walking_speed, walking_steadiness, dietary_energy, dietary_protein, dietary_carbs, dietary_fat
Cloud access (OAuth)
The Cloudflare Worker adds OAuth 2.0 authentication for remote access from Claude.ai or VS Code.
Public URL: https://apple-health-fly-mcp-worker.fjcabello.workers.dev/mcp
VS Code configuration
// .vscode/mcp.json
{
"servers": {
"apple-health-cloud": {
"type": "http",
"url": "https://apple-health-fly-mcp-worker.fjcabello.workers.dev/mcp"
}
}
}
Worker source
See fjcabello/apple-health-fly-mcp-worker (../apple-health-fly-worker/ in the local workspace). Same multi-MCP OAuth gateway pattern as garmin-connect-mcp's mcp-oauth-gateway.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.