Your AI Plan Decides Whether You Can Add an Apple Health MCP Server
You have your export. You have the server. Then step three stops you: your AI plan will not let you add the connector. Here is the eligibility matrix for Claude and ChatGPT plans, the first time anyone has published it, plus the local path that skips the quota completely.
Quick answer
On the hosted-connector path, your AI plan is the wall. Claude Free allows one custom connector at a time, so a single health connector fills the slot. ChatGPT Free and ChatGPT Go allow none, and the ChatGPT phone apps cannot add one either. On ChatGPT Business, Enterprise and Edu, an administrator decides. The way around all of it is to stop treating the health server as a hosted connector: a local MCP server that reads a file on your own machine is not a vendor-hosted connector, so no plan meters it and no admin approves it.
The matrix nobody published
Every vendor documents its own connector and custom-server rules in scattered help pages. We read the setup pages for the two biggest agent platforms on 18 October 2026 and pulled the rules into one table. This is the table a buyer hits at step three of every "connect Apple Health to your AI" guide, and it is the reason a working setup suddenly fails.
| Where you want the server | Can you add it? | The gate |
|---|---|---|
| Claude, Free | Yes, one | One custom connector at a time, so a health connector occupies the only slot |
| Claude, paid tiers | Yes | More connector room; still a hosted connector with a vendor in the path |
| Claude Desktop, local server config | Yes | Not a connector. Read from your own config file, so the connector count does not apply |
| ChatGPT, Free | No | Custom MCP servers are not available on Free |
| ChatGPT, Go | No | Custom MCP servers are not available on Go |
| ChatGPT, Plus and above | Yes | Paid tier required; hosted connector, so the vendor carries the request |
| ChatGPT, phone apps | No | The mobile apps cannot add a custom MCP server at all |
| ChatGPT, Business / Enterprise / Edu | If admin allows | Your workspace administrator decides whether you may add one |
Two consequences fall straight out of that table. First, the cheapest AI plans are exactly the ones that cannot reach your health data through a hosted connector, which is backwards for anyone testing this on a free tier. Second, the moment your setup depends on a connector, your data path runs through the vendor's servers, and someone else's quota sits between you and your own heart rate.
How we sourced this. The plan and connector rules above come from the vendors' own setup pages, read and hashed on 2026-10-18 (evidence ids E2 and E3 in our weekly market pack). We did not test every tier by hand. Plan entitlements change without notice, so confirm yours at the source before you pay for an upgrade to reach a connector.
Why the wall is a connector, not a model
This trips people up. The limit is not the model's ability to read data. Models are fine at that. The limit is the transport: a remote MCP connector is a hosted integration, and hosted integrations are a paid, metered, administrated feature on every big platform.
So the question that decides your setup is not "which model is smartest", it is "where does the server run".
On-device and local vs hosted connector
There are two ways to put your health data in front of an AI. They look similar in a demo and behave nothing alike in practice.
| Local server (on-device data) | Hosted connector (cloud) | |
|---|---|---|
| Where data lives | Your iPhone and your own disk | A vendor server you do not control |
| Plan gate | None. Not a connector | Free tiers blocked, admin-gated on business plans |
| Quota | Unlimited local reads | Metered per request on some rivals |
| Account needed | No account for the data path | Vendor account and plan required |
| Works offline | Yes, the data is already local | No |
| Mobile agents | Any agent that reads a local config | Blocked on the phone apps |
None of that makes a hosted connector a bad product. It is a fine shape for a team that wants zero setup. It is a bad shape for someone who wants their own health data on their own machine with no quota and no administrator. Those are different buyers, and the category is now visibly splitting three ways: file-snapshot local, LAN-served local, and the cloud connector. The cloud connector is the only one of the three charging a recurring meter.
The setup that never hits the matrix
A local MCP server is not a hosted connector, so none of the plan rules above apply to it. MetricBridge ships one: a zero-dependency stdio server that reads your local Apple Health JSON export and exposes 16 read-only tools. It requires Node.js 18 or newer and a path to your export folder. That is the whole dependency list.
{
"mcpServers": {
"health-export": {
"command": "node",
"args": ["/Users/you/.health-export-mcp/server.mjs"],
"env": {
"HEALTH_DATA_DIR": "/Users/you/Library/Mobile Documents/iCloud~ai~healthexport~app/Documents"
}
}
}
}Drop that into a desktop agent's MCP config, quit and relaunch it, then ask the agent to call get_mcp_status. It reports whether the server is up and whether it found your data. If it did, you are reading your own health data through an agent with no connector added and no quota attached. Full per-client recipes live in the Claude setup guide and the multi-client guide.
What you can actually ask
Once a local server is connected, the questions are ordinary ones. How did HRV trend this week against last week. Which nights did my sleep stages change. Has my resting heart rate drifted over the last 30 days. The agent calls get_trends, compare_periods or correlate_metrics and answers in plain language. Everything is a deviation from your own baseline, not a verdict about your health.
FAQ
Can I add an Apple Health MCP server on Claude Free?
A remote custom connector is plan-gated on the Claude apps: Free allows one custom connector at a time, so a single health connector fills the slot. A local MCP server on Claude Desktop is a different path. It is read from your own config file, not added as a connector, so the connector count does not apply to it.
Can I add a custom MCP server to ChatGPT?
Per the setup pages we read on 2026-10-18, ChatGPT Free and ChatGPT Go cannot add custom MCP servers at all, and the ChatGPT phone apps cannot add one either. Custom servers need a paid desktop or web plan, and on Business, Enterprise and Edu an administrator decides whether you may add one. Plan limits change, so check your own plan before you buy anything.
Is there a way to use an Apple Health MCP server with no plan quota?
Yes. A local MCP server that reads a file on your own machine is not a hosted connector, so no vendor plan meters it. MetricBridge ships a zero-dependency stdio server that reads your local Apple Health JSON export. The agent you point at it runs the tool calls on your machine, and the data does not pass through a connector quota or a vendor server.
Do I need a paid AI plan to read my Apple Health data with an AI agent?
Not for the data path. The phone export is local, and a local MCP server reads it locally. What a paid plan buys is the agent capability you want upstream and, on some vendors, the ability to add a hosted connector. If you use a desktop agent that reads a local server config, the plan tier is not the wall.
Doing this yourself? MetricBridge handles the export step: all 190 metrics as local JSON, with the open-source MCP server (GitHub).
One purchase, no connector, no quota
MetricBridge is a $24.99 one-time unlock. It covers the export, the raw MCP server, and the on-device Ask and Weekly Brief. No subscription, no account, nothing on our servers.