How to Keep Your Apple Health Data Private When Using AI Agents
You can ask an AI about your sleep, HRV and heart rate without a copy of your health record landing on someone else's server. Here is which routes keep it on hardware you control, and which do not. (Short version: ChatGPT Health sends your metrics to OpenAI; a local MCP server and an on-device app do not.)
AI-assisted, human reviewed. How we use AI
Quick answer: Your health data stays in your control when the whole chain runs on hardware you own: an app that answers on the iPhone, an export that goes only where you point it, and a local MCP server that reads that file and makes no outbound network calls. Cloud products such as ChatGPT Health or AI Health Export (Gemini) send your metrics to remote servers by design. If privacy is the point, pick the route where nothing has to be transmitted. See exactly where your data goes with MetricBridge.
The cloud problem
Every time you ask a cloud AI about your health data, your metrics leave your phone. Not just the question. The actual numbers. Your resting heart rate. Your HRV readings. Your sleep stages. They travel over the network to a server you do not control, get processed by a model you cannot audit, and the company stores the result.
Most people assume that because the answer comes back in a chat interface, the data stayed local. It did not.
Here is how the main cloud health AI tools worked when their pages were read in August 2026 (Health Auto Export re-read on 27 October 2026). Products change; check each vendor's own page.
ChatGPT Health launched in July 2026. You grant ChatGPT access to Apple Health via a data connector. Your metrics are uploaded to OpenAI servers. The data transits OpenAI's infrastructure; what it is retained for is set by OpenAI's health privacy policy, not by you.
AI Health Export (Gemini) sends your Apple Health data to Google Gemini in the cloud for processing. Every question you ask, every metric you query, goes to a Google server.
Health Auto Export runs its MCP server inside the iPhone app over your local network (HTTP recommended, TCP for Node bridges, Premium tier required, and the app has to stay in the foreground while a client is connected, per its help page, read on 27 October 2026). That is a local design. Its separate export automations can also push data to a cloud folder or a webhook; if you point them there, a copy sits on that third-party server.
There are now at least 10 Apple Health MCP servers on GitHub and npm. Most follow the same pattern: export your data to a shared folder or cloud destination, then expose it through an MCP server. Some require DuckDB. Some require Python. A few require Docker. Most involve a network hop somewhere, so check each one for where the file lands.
The local alternative
A private health AI for your Apple Health data does not have to send the data anywhere. Every step of the pipeline can run on hardware you control.
Here is what a local-first approach looks like.
The iOS app reads HealthKit read-only. A background export writes a compact JSON file to a destination you choose: iCloud Drive (synced by Apple, no app developer in the path), a folder on your Mac, a share on your own LAN, or a webhook you run. Or nothing leaves at all: the export is optional, and Ask works on the phone without it.
The MCP server runs as a single Node.js process on your computer. It makes no outbound network calls. It reads the JSON file from the local folder and exposes its tools (16 of 18 read-only; add_event and remove_event write only an events file beside the export) to any MCP-aware agent. Claude, Cursor, opencode, ChatGPT Desktop with MCP support, any of them. The server has no telemetry, no analytics, no phone-home. It is a small npm package (health-export-mcp 1.8.0) with zero dependencies and 23 prompts.
When you ask a question, the agent calls a tool on your local MCP server. The server reads the data file from disk. The agent sees the result. Nothing passed through a server we operate. If your agent is a cloud AI, the numbers it asks for go to that provider; with a local model they stay on your machine.
This is the architecture of MetricBridge. The iOS app handles the export. The open-source MCP server (health-export-mcp on npm and GitHub) handles the agent bridge. Together they form a pipeline that never sends your health data to us or to a cloud AI of ours; the export goes only where you point it.
Two scenarios, one pipeline
You do not need to choose between the on-device chat and the MCP agent bridge. They work on the same data.
Scenario A: You just want to ask questions on your iPhone. Open the app. Type "how did my HRV trend this month?" The answer is computed on the phone, in one of three on-device ways: Guided (no model), Apple's on-device model, or a ~400 MB model you download once for phones without Apple Intelligence. (A fourth, optional Your Mac mode words the answer with Ollama on your own Mac.) You get a provenance card showing the exact samples, the date range and the calculation. The three on-device ways make no network call, and Your Mac mode talks only to your own Mac. Nothing reaches us or a cloud AI. How each mode works.
Scenario B: You want Claude on your Mac to analyze your sleep patterns over the last year. The same iOS app writes the JSON file to iCloud Drive. The MCP server on your Mac reads it. Claude calls get_health_metrics and get_trends against your local data. The same rule applies: the file sits in your iCloud account, and the server reads it there.
Both scenarios rest on the same principle: the data moves only through paths you control.
About the ad SDK. The App Store version of MetricBridge includes Meta ad attribution SDK. It sends three events (install, trial start, purchase) plus the advertising identifier after you grant App Tracking Transparency. No health data whatsoever reaches that SDK. This is enforced in code. The MCP server remains separate, open source, with no outbound network calls.
What the competitors do (as read in August 2026)
MetricBridge makes one of the products discussed. Competitor facts come from their own pages, dated where shown.
AI Health Export (Gemini) is the most direct cloud competitor to MetricBridge. It lets you chat with your Apple Health data but uses Google Gemini in the cloud. Every answer requires a round trip to Google servers. Its page describes no on-device inference and no MCP bridge. Your data leaves your phone.
Ask My Health is a newer iOS app with a built-in MCP server. It processes 100 free analyses on device, then needs a monthly or yearly Pro subscription. The app claims no data collection and covers workouts, heart rate, HRV, nutrition, training load and recovery. Its listing does not state a metric count, provenance cards or a weekly digest.
HealthMirror syncs your Apple Health data to iCloud Drive as JSONL files and provides its own MCP server. It is free and depends on iCloud sync. Its page describes no AI chat and no on-device question answering.
The on-device health AI apps (Risi, Metrya, Sifa, Thryve, Lifetrails, Ahimo, Suvi, Pharo, Priv AI) process data locally. When their pages were read, none documented an MCP bridge, so getting that data into Claude or Cursor needs a separate export tool. (Marrow, covered in its own comparison, does ship one.)
MetricBridge gives you both: answers on the phone with the working shown, and a zero-dependency MCP pipeline to your own agents. Same data, two surfaces. Every number links to the published method.
Who this is for
The developer running Cursor who wants to query their HRV against their workout log. The quantified-self person who wants Claude to run weekly trend analysis. The privacy-conscious iPhone user who will not paste health data into a cloud chatbot. Anyone who has looked at ChatGPT Health and wondered "wait, does that send my heart rate to OpenAI?"
The answer is yes. ChatGPT Health sends your data to OpenAI. Every product that processes in the cloud does the same. The surest way to keep your Apple Health data in your control is to not send it anywhere.
That is what a local-first, on-device setup does. No cloud. No upload. No server in the path. Wellness information, not medical advice: where a number moves, the app says "a deviation from your baseline", never a diagnosis. For the step-by-step setup, see Connect Apple Health to Claude via MCP. To understand the export layer, see Export Apple Health data to JSON for AI.
Frequently asked questions
Does ChatGPT Health keep my Apple Health data private?
ChatGPT Health sends your Apple Health data to OpenAI servers for processing. Your metrics leave your device and transit their infrastructure. It is a cloud product, so the data necessarily travels over the network.
What is the most private way to use AI with my Apple Health data?
Use a local-first toolchain: an app that answers questions on the device and an MCP server that makes no outbound network calls. MetricBridge plus its open-source MCP server meets this standard today. The privacy page lists every data flow.
Do any on-device health AI apps export to Claude or Cursor?
Few do. Of the apps read for this post (Risi, Metrya, Sifa, Thryve, Lifetrails, Ahimo, Suvi, Pharo, Priv AI), none documented an MCP bridge; Marrow does. MetricBridge answers on the phone and ships an MCP bridge.
Is MetricBridge really private?
The on-device chat processes everything on your iPhone. The MCP server makes no outbound network calls. The iOS app does include Meta's ad attribution SDK that sends three events (install, trial start, purchase) plus the advertising identifier after you allow tracking, but no health data ever reaches that SDK. Your health metrics, questions and answers never reach us or a cloud AI.
Can I use MetricBridge without the ad tracking?
Yes. iOS shows an App Tracking Transparency prompt on first launch. If you tap Ask App Not to Track, the ad SDK sends nothing. The health features work identically either way.
Health AI that never reaches us or a cloud AI
Ask your Apple Health data a question and see the days, the samples and the calculation behind the answer. Your health data never touches a server of ours, and the MCP server makes no outbound network calls.
Where it goes: answers are computed on your iPhone; the export goes only where you point it. What you get: Ask with a provenance card under every answer, the Weekly Brief, 194 metrics as JSON, the MCP server. Setup: about two minutes. $24.99 one time after a 7-day free trial, not a subscription. Works on any iPhone; a ~400 MB download covers phones without Apple Intelligence.