ChatGPT Health vs On-Device Health AI: What Actually Leaves Your iPhone

Quick answer: with ChatGPT Health, the Apple Health categories you share are processed on OpenAI's servers. With an on-device health AI, every answer is computed on your iPhone and nothing reaches a cloud AI. You end up with the same kind of answer either way; the difference is where your heart rate, sleep stages and HRV go to get it.

By MetricBridge · Published 18 August 2026 · Updated 27 October 2026 · ~11 min read

AI-assisted, human reviewed. How we use AI

Disclosure. MetricBridge makes one of the products compared. Third-party facts come from the vendors' own pages, read on 27 October 2026. Updated 27 October 2026 to match OpenAI's current Health pages.

On 23 July 2026, OpenAI began rolling out ChatGPT Health to logged-in US users aged 18 and over, on web and iOS. It connects Apple Health, plus medical records from supported US provider portals, One Medical and Function Health (press coverage also named Epic and Oracle Health systems), so ChatGPT can answer questions about your health data. By OpenAI's own figure, more than 300 million people a week turn to ChatGPT with health-related questions.

AppleInsider, TechCrunch and The Verge all covered the launch, AppleInsider specifically on privacy. The honest headline: the heart rate, sleep stages, blood oxygen readings and records you connect are processed on OpenAI's servers.

This post compares ChatGPT Health with on-device health AI apps: exactly what leaves your iPhone with each, and what stays on it.

How ChatGPT Health works (and where your data goes)

ChatGPT Health is a cloud service. Here is the data flow:

  1. You connect Apple Health in the ChatGPT iOS app. The standard iOS Health permission screen lets you pick which categories to share.
  2. The categories you picked sync to OpenAI's servers: heart rate, HRV, sleep stages, steps, workouts, blood oxygen, and any other HealthKit types you enabled.
  3. OpenAI's models answer in the cloud. OpenAI names GPT-5.5 Instant for free users and GPT-5.6 Sol for paid plans as its Health models (openai.com, 23 July 2026). By default ChatGPT asks before using your Health data in a conversation.
  4. Training and retention, in OpenAI's own words. Connected Apple Health and medical-record data, and the conversations that use it, are not used to train OpenAI's models regardless of your training setting, and synced data is deleted within 30 days of disconnecting the source. Handling is covered by the Health Privacy Notice (updated 29 June 2026), which also addresses Washington's My Health My Data Act and Nevada's consumer health data law.

Those are better terms than many cloud products offer. The core fact still stands: your resting heart rate and sleep architecture are processed and held on a third party's servers while the source is connected. Whether that is fine is your call.

How on-device health AI works (and what stays on your iPhone)

On-device health AI takes a different approach. MetricBridge answers in the app on the phone; Risi.ai runs local models with Apple's MLX framework; Ask My Health runs a local MCP server plus Siri Intents on the phone rather than an in-app chat (100 free analyses, then a Pro subscription, per its App Store listing). For MetricBridge the flow is:

  1. You install the app and grant Apple Health read-only access.
  2. You ask a question in plain language, like "how is my HRV trending this month?"
  3. The app computes the answer on your iPhone. Ask runs four ways: Guided (no model), Apple's on-device model (Apple Intelligence), a downloaded model of about 400 MB for phones without Apple Intelligence, or, if you turn it on, Ollama on your own Mac wording the answer over your Wi-Fi. Deterministic code computes the numbers; the model only narrates them. Where your data goes when you ask shows each mode.
  4. Every answer renders as a provenance card showing the headline number, the metric, the date range, and exactly which samples produced it.

No account. No sign-in. No health data reaches us or a cloud AI; the export, if you turn it on, goes only where you point it. For MetricBridge, the open-source MCP server makes no outbound network calls. The app's Meta advertising SDK (which sends install, trial start and purchase events plus the advertising identifier if you grant App Tracking Transparency) cannot access HealthKit data: enforced in code with a CI test that fails the build if any HealthKit import reaches the SDK wrapper.

Side by side: data flow comparison

Dimension ChatGPT Health On-Device Health AI
Where queries are processed OpenAI cloud servers Your iPhone
Health data reaches a cloud AI Yes, OpenAI's servers No
Who holds your raw metrics OpenAI, while connected (deleted within 30 days of disconnecting) Only you
Account required OpenAI account None
Model capability GPT-5.5 Instant (free), GPT-5.6 Sol (paid), in the cloud Guided (no model), Apple's on-device model, a ~400 MB downloaded model, or Ollama on your own Mac
Used to train the vendor's models No, per OpenAI, regardless of setting No model of ours ever sees it
Medical records (provider portals) Yes No
Provenance cards (shows the numbers) No Yes
MCP bridge for AI agents Not stated by OpenAI Yes (MetricBridge and a few others)
Weekly automated digest No Yes (MetricBridge)
Works with Claude, Cursor, opencode Not stated Yes (via MCP server)

The MCP bridge difference

On-device health AI apps are private by design, but most are walled gardens. You ask questions in the app, you get answers there. Your data never reaches a cloud AI, but you also cannot feed it to the AI agents you already use.

If you use an on-device health AI that also ships an MCP bridge, you get both: the privacy of on-device processing and the ability to query your health data from Claude, Cursor, ChatGPT Desktop, or any MCP-aware agent. MetricBridge provides both, using a zero-dependency Node.js MCP server that makes no outbound network calls, with no app to keep open on your computer and no server of yours. It exposes 21 tools, 19 of them read-only (see the MCP docs) including get_health_metrics, get_trends, and compare_periods.

Real example: same question, two architectures

Consider the question: "How did my HRV trend this week compared to last week?"

ChatGPT Health

Your question and the Apple Health categories you shared go to OpenAI's servers. The model does the comparison and returns an answer. Your HRV data sits on OpenAI's infrastructure while the source is connected.

MetricBridge (on-device + MCP)

Via the app or through Claude over MCP, the on-device parser reads your local JSON cache, computes week-over-week trend math on your iPhone (in the app) or on your own computer (over MCP), and returns: "HRV averaged 61 ms this week, up 4.4% from 58 ms last week. Date range: 11 Aug to 17 Aug vs 4 Aug to 10 Aug. Sample count: 1,247 readings." The numbers come from code you can check in the published methods. No health data reaches us; over MCP, only the numbers your agent asks for go to its model provider.

Where ChatGPT Health is stronger

On-device health AI is not a replacement for ChatGPT Health in every dimension:

  • Model size. A model that fits on an iPhone is small. OpenAI's largest models are better at long, open-ended reasoning and summarising literature.
  • Medical records. ChatGPT Health connects supported US provider portals, One Medical and Function Health. On-device wellness apps do not, and MetricBridge keeps that boundary on purpose: interpreting clinical records crosses the wellness-to-medical line. Wellness information, not medical advice.
  • Zero setup. If you already use ChatGPT, connecting Health is a few taps. On-device apps are a separate install.
  • Cross-app integration. ChatGPT Health can use data from apps like MyFitnessPal that write to Apple Health. On-device AI reads HealthKit too, so the same written-through data is there, but only what reaches HealthKit.

For the core use case of asking questions about your daily Apple Health metrics and getting answers you can check, on-device AI is comparable day to day and better on where the data goes.

Which should you use?

Choose ChatGPT Health if: you want OpenAI's largest models, need to connect medical records from provider portals, and are comfortable with OpenAI processing your health data under its Health Privacy Notice.

Choose an on-device health AI if: you want answers computed on your phone that never reach a cloud AI, want every answer to show the exact numbers behind it, and want to use your data with Claude, Cursor or other AI agents without creating another account.

You can also use both. They are not mutually exclusive.

Frequently asked questions

Does ChatGPT Health send my Apple Health data to OpenAI servers?

Yes. The categories you pick on the standard iOS Health permission screen sync to OpenAI's servers, where the model reads them. OpenAI says this connected health data is not used to train its models regardless of your training setting, is deleted within 30 days of disconnecting, and that ChatGPT asks before using it by default (Health Privacy Notice, updated 29 June 2026; read 27 October 2026).

What data does an on-device health AI keep on my iPhone?

An on-device health AI like MetricBridge computes every answer on your iPhone: Guided mode with no model, Apple's on-device model, a downloaded model of about 400 MB for phones without Apple Intelligence, or, if you turn it on, Ollama on your own Mac wording the answer. No health data reaches the developer or a cloud AI. The app reads Apple Health read-only and renders every answer as a provenance card showing which samples and date ranges produced it.

Can I still use my Apple Health data with Claude or other AI agents if I use an on-device app?

Yes, if the on-device app also ships an MCP bridge. MetricBridge provides a zero-dependency MCP server (open source, npm, no outbound network calls) that lets Claude, Cursor, ChatGPT Desktop, and other MCP clients query your Apple Health data locally. Export and raw MCP access are a one-time unlock, never metered.

Which is more private: ChatGPT Health or an on-device health AI?

An on-device health AI is more private because no health data reaches a cloud AI; the computation happens on your phone. ChatGPT Health processes your data on OpenAI's servers under its Health Privacy Notice. The trade-off is model size: ChatGPT Health uses OpenAI's largest models, while on-device AI uses a model that fits on the iPhone.

Does MetricBridge send any data anywhere?

No health data reaches us or a cloud AI. The export, if you turn it on, goes only where you point it (iCloud Drive, a folder, your LAN or a webhook). The app's Meta advertising SDK sends install and purchase events if you grant tracking permission, but no HealthKit data can reach it: enforced in code with a CI test. The MCP server makes no outbound network calls.

Get answers that stay on your phone

Ask your Apple Health data anything, see the numbers behind every answer, and let your own agents read the same data. 194 metrics, 21 MCP tools (19 read-only), no account.

Where it goes: answers are computed on your iPhone; no health data reaches us or a cloud AI. The export goes only where you point it. Setup takes about two minutes. $24.99 one time after the trial, everything included, never a subscription. Works on any iPhone: a ~400 MB downloaded model covers phones without Apple Intelligence. Wellness information, not medical advice.