Best MCP Servers for Health Data in 2026: A Buyer's Guide
You want your agent to answer from your real health data, read-only, with nothing to host. MCP is the standard that makes that happen. Here is the field of servers, the four things that matter when you pick one, and the one design that gives the privacy away.
MetricBridge makes one of the products discussed. Competitor facts come from their own pages, dated where shown.
AI-assisted, human reviewed. How we use AI
Quick answer. If your data lives in Apple Health, run health-export-mcp: a zero-dependency server that reads 194 metrics through 18 tools (16 read-only), fed by an iOS app that exports on a schedule, so you never hand-export XML. If you need Oura, Whoop or Garmin data alongside it, look at the per-wearable connectors in the Delx Wellness registry or a self-hosted server such as FreeReps. Everything below explains why, and what to watch for.
What an MCP server for health data actually does
The Model Context Protocol (MCP) is the open standard that lets an AI agent call external tools through one interface. A health-data MCP server is a translator: it reads your health measurements from somewhere (a JSON cache, a SQLite file, a CSV dump) and advertises query tools like get_health_metrics and get_trends that any MCP-aware client can call.
The useful consequence: your agent never needs a copy of your numbers pasted into the chat. It asks the server for the data on demand, then reasons over live values. No stale CSV, no mangled units.
Four things that actually matter when you pick one
- Where the data goes. A local server reads a file on your machine and makes no outbound network calls. A cloud-hosted server means your heart-rate history transits someone else's infrastructure on every query. For health data, that is the difference between private and not.
- Dependency weight. Zero-dependency servers start in seconds and install with nothing to compile. Servers built on DuckDB or Python pull native bindings and transitive packages, and sometimes require a separate third-party app just to produce their input file.
- Tool and metric coverage. More read-only tools and more metrics means your agent can actually answer the question instead of hitting a wall at "workout data only" or "HRV only."
- How the data gets there. The best setup writes your data automatically from an iOS app. The worst makes you export a multi-gigabyte XML file by hand every month.
The Apple Health MCP servers, side by side
There are more than a dozen Apple Health MCP servers on GitHub (search read 27 October 2026). Most are readers for an export file: they need DuckDB or a set of Python packages, and they ship no iOS app, so you produce the file yourself. A few do ship an app (Health Auto Export, FreeReps, Open Wearables, and smaller projects such as roian6/apple-health-ai-bridge and jefflitt1/health4ai). The table shows the ones worth knowing, with third-party facts read on 27 October 2026.
| Server | Dependencies | Coverage | iOS app |
|---|---|---|---|
health-export-mcp |
Zero (84 kB npm package) | 194 metrics, 18 tools (16 read-only) | Yes (MetricBridge) |
| the-momentum | Python, Docker or uv; DuckDB, Elasticsearch or ClickHouse backend | Apple Health via manual export.xml | No (README says the project has evolved into Open Wearables) |
| neiltron/apple-health-mcp | Node 22+, duckdb npm package; CSV from the Simple Health Export CSV app | Apple Health, 3 tools | No |
| rinoshiyo/apple-health-mcp-server | 9 runtime PyPI packages (DuckDB-based), v0.6.1 | Apple Health | No |
| Health Auto Export | Built into the iOS app: Streamable HTTP on your LAN (bearer token); TCP is the legacy transport; an official Node bridge (8 tools, MIT) speaks stdio | Apple Health; Premium tier required, app must stay in the foreground | Yes (iPhone, iPad, Mac, Vision) |
The dependency column is the one to stare at. health-export-mcp (1.8.0 on npm) has zero dependencies. The DuckDB path pulls a C++ engine plus its bindings. The neiltron server reads a CSV produced by a separate app; its README says Apple's native export.xml is "not currently supported" (read 27 October 2026). That is one more app and one more manual step before the first query. Health Auto Export keeps everything on your LAN, but the phone has to stay awake in the foreground while an agent is connected; its help page is clear about that.
Wearable data: Oura, Whoop, Garmin
If your data lives outside Apple Health, the field is thinner (read 27 October 2026). Delx Wellness is a registry of 16 separate connectors, each its own npm package: WHOOP, Oura, Garmin, Strava, Fitbit, Withings and Apple Health among them. Its davidmosiah/apple-health-mcp package reads an export.xml or zip, not live HealthKit. FreeReps is a self-hosted health server (Docker Compose, Tailscale required) with MCP over stdio and Streamable HTTP and 20 tools; its sources are Apple Health, Oura, Withings, Hevy and Alpha Progression, not Whoop or Garmin. It has an iOS companion app. Both are a fit if you want one stack across devices and are happy to run it.
The cloud trap
At least one self-deployed example exists of an MCP server hosted off the device, a personal Fly.io deployment that syncs an export through a Cloudflare Worker. Understand what that design means before you copy it: on every query, your health data crosses a server you do not run. If privacy is the point, the server should run on your machine and read a file you control. Where MetricBridge's data goes, mode by mode.
Which one should you run?
If you are an Apple Health user who wants an agent to answer real questions (sleep trends, HRV recovery, resting-heart-rate changes), take health-export-mcp plus the MetricBridge iOS app. The app exports your 194 metrics read-only to a folder you pick (iCloud Drive, a synced folder, your LAN, or a webhook you run), and the server just reads that folder. Nothing transits the developer's servers, and the model never sees a cloud round-trip.
If you are all-in on Oura or Withings and don't mind running your own stack, FreeReps is the move; for Whoop or Garmin, the Delx Wellness connectors. If a tool's pitch starts with "cloud-hosted" and your data is health data, ask where the file lands before you sign up.
Setting it up yourself? MetricBridge is the on-device health AI that also ships the open-source MCP server. It covers all 194 metrics, 18 tools (16 read-only), and the whole thing is a single zero-dependency npm package. A one-time $24.99 purchase covers export, MCP, and the on-device chat. No subscription, nothing metered.
{
"mcpServers": {
"health-export": {
"command": "npx",
"args": ["-y", "health-export-mcp"],
"env": { "HEALTH_DATA_DIR": "~/Library/Mobile Documents/iCloud~ai~healthexport~app/Documents" }
}
}
}Your agent, answering from your real data
MetricBridge exports 194 Apple Health metrics as clean JSON on a schedule, and health-export-mcp serves 18 tools (16 read-only) over MCP. Nothing to host, nothing on our servers.
Where does it go: the export lands in iCloud Drive, a folder, your LAN or a webhook you run; the server reads that file and makes no outbound network calls. Setup: one config block, about two minutes. $24.99 one time after a 7-day free trial, not a subscription; the server is free and open source. Works with Claude, Cursor, opencode, ChatGPT Desktop and any MCP client.
Frequently asked questions
What is the best MCP server for Apple Health data?
health-export-mcp is a strong default. It's a zero-dependency Apple Health MCP server, it exposes 18 tools (16 read-only) across 194 metrics, and its iOS app exports on a schedule, so you're not hand-exporting XML.
Do any MCP servers for health data work with Oura, Whoop, or Garmin?
Yes. Delx Wellness publishes separate connectors for WHOOP, Oura, Garmin and others. FreeReps is a self-hosted server covering Apple Health, Oura and Withings (not Whoop or Garmin) if you're willing to run your own hardware. Read 27 October 2026.
Is a zero-dependency MCP server actually better for health data?
For health data, yes. A zero-dependency server has no native bindings to compile, a smaller install surface, no transitive package risk, and it starts in seconds. DuckDB-based servers pull native extensions and some need a separate CSV export app to produce their input.
Does a cloud-hosted MCP server keep my health data private?
No. A cloud-hosted MCP server means your health data transits a third-party server on every query, which defeats the point of on-device health AI. Prefer a local-only server that reads a file on your own machine and makes no outbound network calls.