On-Device Health AI That Actually Answers Questions: No Cloud, No Upload
Ask "how's my sleep trending this month?" and get a real answer, with the days, the samples and the calculation under it, computed on your iPhone. No upload, no account. Here is what that looks like and how it compares to the apps that send your data to a server.
AI-assisted, human reviewed. How we use AI
You've got years of Apple Health data on your phone. Heart rate, sleep stages, HRV, resting pulse, step counts, workouts: 194 metrics, with readings going back months or years.
What can you actually do with it?
Most people open the Health app, scroll through a few charts, and close it. The data is there, but it's locked inside a grid of screens. You can't ask it a question. You can't say "how's my sleep trending this month?" and get a real answer.
That's the gap an on-device health AI fills. And the privacy part isn't optional. It's the whole point.
What "on-device" actually means for your health data
Here's the problem with cloud-based health AI: you have to upload your data to use it. Every single metric (your resting heart rate, your sleep patterns, your HRV) gets copied to someone else's server. The company promises to "protect your privacy" in a blog post, but your data still leaves your phone.
On-device processing flips that. The AI model runs on your iPhone. Your health data never reaches us or a cloud AI. The question, the computation, the answer: all on the phone. No upload, no cloud account, no server in the path.
That's what MetricBridge does. It reads Apple Health read-only, answers questions about your data on-device, and shows you exactly how it got each answer. You can also export to your own AI agents over MCP, but the core experience never needs to touch the internet.
Real questions you can ask
The chat is a text field at the bottom of the screen. You type a question in plain English and get an answer in seconds. It runs on the phone in three live ways: a deterministic Guided engine (every iPhone, no model), Apple's on-device model on phones with Apple Intelligence, or a ~400 MB model you download once on phones without it. What each mode sends, side by side.
Try these:
- "How's my sleep trending this week vs last?" It pulls your sleep samples for both periods, computes the average, and shows the delta. You see the numbers: "7.2h this week vs 6.8h last week (+5.9%)."
- "Compare my HRV this month to last month." HRV is a lagging recovery signal. The app computes the mean, the trend direction, and the variance. If your HRV dropped 12% month-over-month, you see it in a chart with the exact values.
- "What's my resting heart rate been doing on weekends?" It filters weekend vs weekday readings, computes the average, and tells you whether your weekends are actually more relaxed.
- "Show me my step count for the last 14 days." Returns a clean list or chart with the date, step count, and daily average.
Every answer comes with a provenance card: the headline number in large type, an inline chart, and a tappable strip that shows exactly which samples, date range, and computation produced the result. No "AI vibes." You see the data that drove the answer.
Why provenance matters more than you think
This is where most health AI apps fall apart. They return a nice-sounding sentence ("Your sleep quality has improved this week!") but you can't see the numbers behind it. Which metric? Compared to what baseline? Show me the actual data.
MetricBridge's provenance card is the answer to that. Every computed result shows:
- The metric and date range queried
- The raw sample count and computation (mean, median, trend, delta)
- A chart or sparkline of the underlying data
- A tappable strip that expands to individual samples
You can verify the AI's work. If it says your HRV averaged 62 ms this week, you can tap through and see the individual readings (58, 64, 67, 61, 59) and confirm the math. It's the difference between trusting an app and knowing you're right. Each calculation is published at /methods/.
Why provenance matters. A score that moves with no reason given is the cost of a black box: "recovery good" one day, "recovery poor" the next, and no way to tell which is right. Provenance makes the answer auditable. Where a number moves, the app says "a deviation from your baseline", never a verdict. Wellness information, not medical advice.
Also: the Weekly Health Brief
You don't have to open the app to get value. The Weekly Brief is an on-device digest that arrives as a notification + in-app card. It summarizes your week's trends (sleep, HRV, resting HR, steps, workout load) and compares them to the previous week. Same provenance cards. Same local computation. It's the renewal engine: the app shows up for you.
No server generates the brief. No cloud processes your data. Just your iPhone, running the same numbers you'd compute yourself if you had the time. A Today card with a readiness estimate and its range is built and arrives in the next free update; what it will show.
The other side: export to your own AI agents
If you're the kind of person who runs Claude, Cursor, or opencode on your own machine, the on-device chat isn't the only way in. MetricBridge also exports your 194 metrics as clean JSON to a destination you choose (iCloud Drive, a local folder, your LAN, or a webhook) and ships a zero-dependency MCP server (16 read-only tools) that lets any MCP-aware agent query your live data.
That means you can ask Claude the same questions about your sleep and HRV, running on your Mac, against the same local data. The MCP server is read-only, open-source, and makes zero network calls. It's the developer counterpart to the on-device consumer app.
For the full setup, see Connect Apple Health to Claude via MCP.
What about the other on-device health apps?
There are a few apps in this space now: Risi.ai, Metrya, eutonia and Sifa all describe some form of on-device processing (as read on their pages in July 2026). None of those documented an MCP bridge or a provenance card on every answer. For dated, side-by-side detail see on-device vs cloud health AI apps and the comparisons hub.
MetricBridge gives you both: a private on-device chat with auditable answers, and a live data pipeline to your own agents. You don't have to choose between privacy and utility.
Pricing and what you get
The app is straightforward. 7-day free trial, then $24.99 one time. Everything included. No subscription. That covers export, MCP access and the on-device intelligence: Ask with provenance cards, the Weekly Brief and live agent tools. It is permanent and never metered. No accounts, and your health data never reaches our servers. The app does report installs and purchases to our advertising partner (only if you allow tracking when iOS asks) and never any health data.
You can start with the free trial, grant HealthKit read access, and see the data flow before committing. The MCP server itself is free and open source.
Your data. Your phone. Your answers.
Ask real questions about your Apple Health data, answered on your iPhone, with numbers you can check.
Where does it go: 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.