Health Auto Export ships an on-device MCP server, moving MCP into consumer iOS apps

An MCP server in an App Store listing

Health Auto Export, an iOS app for exporting Apple Health data, now advertises an integrated MCP server in its App Store listing. According to a teardown published on dev.to, the claim is accurate — but the label covers one particular architecture with real constraints, and it is one of the clearer signs yet that the Model Context Protocol is moving from developer tooling into mainstream consumer apps.

MCP is an open protocol that lets AI clients such as chatbots and coding agents discover and call external tools and data sources. It has spread quickly through developer tooling; a health app shipping a server as an advertised feature points at a different, consumer-facing stage of adoption.

How the on-device server works

Per the vendor's help pages, as quoted in the dev.to write-up, the server runs on the iPhone itself. AI clients on the same local network connect to an HTTP endpoint on port 9000, with Streamable MCP as the recommended transport, and authenticate using a bearer token. A TCP option is available for Node-bridge setups.

The catch is that the app has to stay in the foreground while clients are connected. Data stays on the local network, but an agent can only reach it while the app is open, the phone is awake, and both devices share a Wi-Fi network. For an assistant meant to answer health questions on demand, that foreground requirement is the practical failure mode.

The alternative: snapshots and a local process

The teardown contrasts this with a stdio-based server: a plain Node process the agent launches over standard input and output, with no sockets, ports or tokens exposed on the network. In that arrangement, the iOS app writes a JSON snapshot to a destination the user picks — iCloud Drive, for instance, which syncs it to a Mac — and the server reads that folder whether or not the phone is awake.

Because a snapshot is not live by definition, the design surfaces staleness explicitly. According to the write-up, exports are flagged as stale after 26 hours by default, and responses include metadata such as age in hours, an as-of timestamp and the last data date, so an agent can say how old an answer is instead of quietly serving yesterday's numbers. That server ships 16 read-only tools and is MIT-licensed, the write-up notes.

A correctness wrinkle

The teardown also raises a data-quality issue it attributes to an open issue on the incumbent app's tracker, number 56: when another app backfills historical samples, they are stamped with a recent HealthKit write time, so an export built on changes since the last sync never recomputes those earlier days. The author says a competing tool, MetricBridge, addressed this in its 1.10 release by recomputing the day a backfilled sample belongs to. Worth noting: the write-up links to healthexport.dev and covers that rival tool in favourable terms, so the comparison comes from an interested party.

Why it matters

A consumer health app treating MCP as a headline App Store feature signals the protocol's shift from developer infrastructure to a mainstream integration surface. Health data is among the most sensitive material on a phone, and both architectures described here keep it local, with no cloud intermediary — plausibly the template for consumer AI integrations going forward. The story is also a lesson in reading claims carefully: "has MCP" can mean a live server tied to a phone that must stay open, or a file-based bridge that trades freshness for availability. Asking where the server runs, how the agent reaches it, and what happens when the phone locks tells you which one you are getting.