Health data × AI

Put your health data to work with AI.

Choose a file upload, the bundled Mac MCP helper, or the standalone CLI/MCP preview. Each gives an assistant a different path to selected Apple Health data.

These local-first paths do not create a Health.md server copy by default. Agent hosts and model providers may process the files or returned values you share under their own policies.

Choose your workflow

Files, Mac context, or direct-phone preview

1. Upload a selected file

Export selected dates and metrics as CSV or JSON on iPhone, then attach the files to an assistant that supports them. This is a snapshot; later exports need another upload. Follow the Claude guide or ChatGPT guide to inspect source values before asking for trends.

2. Query the bundled Mac context

The released Mac app includes a signed healthmd-mcp helper. Keep Health.md for Mac open. Queries can use previously acquired encrypted context without a live iPhone connection; a fresh acquisition needs the authorized, reachable iPhone and the required foreground state. Check freshness and scope before answering.

Connect the bundled Mac helper

3. Use the standalone preview

The separate healthmd CLI provides healthmd mcp serve on macOS, Linux, and Windows without the Health.md Mac app. It pairs directly with a compatible iPhone; keep Health.md foreground on the phone for fresh typed queries. This remains an explicitly unqualified preview, not a stable 1.0 release.

Start with the Apple Health and Health Connect CLI first-export guide, then follow the standalone query guide. Check release status and compatibility before installing.

Choose the connection before copying an MCP configuration. The bundled helper and standalone CLI have separate pairing, credentials, context, and jobs. A successful readiness check is not release or OS qualification. For Android files, start with the Health Connect export guide; the documented standalone typed-query preview requires iPhone.

Per assistant

Setup notes

01

Claude

For a file upload, use the Claude walkthrough. For MCP, choose the bundled Mac or standalone preview instructions above and configure that surface's executable. Inline chart support depends on the host and available tools; inspect capabilities rather than assuming every client renders the same result.

02

Codex

Use the bundled Mac setup for encrypted-context queries, or the standalone preview guide for direct-phone access. Keep their configuration and pairing steps separate. Run readiness and capability discovery before selecting a query.

03

ChatGPT

The ChatGPT file-upload guide starts with a synthetic CSV sample and source-row checks. That tutorial supplies a manual snapshot; it does not configure a live device or MCP connection. This describes the tutorial's path, not every integration a host may support.

Verify before interpreting

Check readiness, freshness, and complete coverage

  1. For MCP, call healthmd_doctor and inspect healthmd_capabilities. Read their checks and next actions; ready does not mean every historical date is authorized or cached data is current.
  2. Use explicit dates, metrics, sources, and detail. For Mac context, inspect the acquired scope and refresh status. For the direct preview, verify the paired phone and its reported query support.
  3. Preserve units, source evidence, recorded-day counts, missing intervals, and limitations. An empty or failed interval is not zero.
  4. Check remaining cursors and page/byte ceilings. A successful request can still leave additional pages or partial coverage.
  5. For uploaded files, check the filenames and source rows used. For every path, review calculations and charts before relying on the answer.

Follow approval prompts for acquisition and writes. Keep a durable job identifier after an accepted refresh, and inspect its status before retrying a timeout. The bundled Mac guide explains partial results and refresh recovery; the compatibility reference distinguishes operational readiness from release qualification.

Privacy

Scoped by default

Choose an explicit scope and inspect what came back. File uploads share selected snapshots. MCP queries return bounded results with coverage, missing intervals, and limitations; preserve those details in the answer rather than treating an assistant's summary as proof of completeness.

Acquisition and file-writing actions have their own authorization and host approval controls. Health.md reports factual observations with units, provenance, coverage, and missingness. It does not diagnose or recommend treatment. Direct device queries do not automatically upload results to Health.md.

The separate, optional single-owner Health.md Cloud pilot retains API exports intentionally uploaded to it. A separately authorized read-only cloud MCP endpoint can return retained fields to an agent. There is no public signup and no automatic device sync. Cloud MCP is distinct from both local MCP paths described above. Review the privacy policy for retention, deletion, and the pilot's unbacked storage limits; your agent or model provider's policies govern its copy.

Keep going

From data to answers

Get the file first

CSV, JSON, Markdown, or Obsidian — export exactly what you want the AI to see. Export Apple Health data

Just need CSV?

The spreadsheet-ready path for Excel, Google Sheets, and AI uploads. Apple Health to CSV

Developer docs

MCP tools, agent architecture, typed queries, and contracts. Local agents documentation

Common questions

Frequently asked questions

Which AI assistants work with Health.md?

Claude, Codex, and compatible MCP hosts can use the chosen Health.md MCP surface. Assistants that accept CSV or JSON can analyze files you provide. Available tools and inline rendering depend on the host; inspect capabilities before choosing a workflow.

Do I need the Mac app to use MCP?

The bundled healthmd-mcp helper requires Health.md for Mac open; existing encrypted context can be queried without a live iPhone, while fresh acquisition needs the connected source. The separate healthmd mcp serve preview does not require the Mac app and pairs with a compatible foreground iPhone. It remains unqualified for stable 1.0. File uploads require neither MCP setup.

Does the agent see all of my health data?

File uploads provide the selected snapshot. MCP queries use explicit dates and metrics and return bounded results with coverage and limitations. Check source authorization, acquired scope, pagination, and freshness; tool success does not prove complete history or complete traversal.

Can Health.md or the AI diagnose me?

No. Health.md reports factual observations with units, provenance, and coverage. It does not diagnose, recommend treatment, or provide medical advice.

Is my health data sent to Health.md's servers for AI use?

Not by default. Local file exports and direct device MCP queries do not create a Health.md server copy. The separate, optional single-owner Health.md Cloud pilot retains API exports intentionally uploaded to it; a separately authorized read-only cloud MCP endpoint can return retained fields to your agent. There is no public signup and no automatic device sync. Agent hosts and model providers may retain or process files and returned values under their own policies.