google-health-mcp

MCP server for the Google Health API, with a local SQLite cache and trend analysis.

Designed for Claude Code and other MCP clients. Your data syncs to a database on your own machine, so queries are fast, work offline, and cost no API quota.

Features

• Local SQLite cache - sync once, query instantly
• Incremental sync - each run fetches only what is new, resuming from where the last one stopped
• Offline mode - serve the cache with no credentials and no network at all
• Trends - weekly, monthly or quarterly aggregates, and two-period comparisons
• ECG - readings stored whole, waveform included, returned only when asked for
• doctor - diagnoses a setup offline and read-only, without spending quota

Data types

Requirements

• Python 3.13+ (tested on 3.13 and 3.14, on Linux, macOS and Windows, in CI)
• A Google account with health data, and a Google Cloud project to authorise against. No billing account is needed - the console offers a free trial throughout setup and you can decline all of it.

Setup

1. Install

pip install google-health-mcp

Or run it without installing, in which case every google-health-mcp ... command you run below becomes uvx google-health-mcp ...:

uvx google-health-mcp --version

2. Create the Google Cloud project

Every user registers their own OAuth client. This is seven console steps, and the page names are Google's as of August 2026.

Google's own setup page will send you somewhere else - follow the steps below instead. Its quick-start builds a Web client with https://www.google.com as the redirect URI, which suits the OAuth Playground rather than a program running on your machine; this server refuses that file and says so. Use that page only to check whether one of the pages below has been renamed.

1. Project. Create a project at console.cloud.google.com/projectcreate and select it.
2. API. Enable Google Health API on the API Enablement page.
3. Get started. Open Google Auth Platform and complete Get started - app name, support email, External audience, contact email. A new project has no Audience, Data Access or Clients page until this is done.
4. Audience. Under Test users, add your own Google account. Skipping this fails sign-in with 403: access_denied.
5. Data Access. Click Add or remove scopes, search for "Google Health API", and tick the read-only scopes listed under OAuth scopes below.
6. Clients. Create an OAuth client of type Desktop app and download its JSON. A Desktop client permits the loopback redirect automatically, so there is nothing to register; a Web client does not, and fails at consent instead.
7. Publish. Back on the Audience page, click Publish app.

Step 7 is the one that bites, and it is worth checking rather than assuming. While an app's publishing status is Testing, Google issues refresh tokens that expire seven days after consent - so everything works, and then syncing stops a week later with nothing pointing back to this moment. The Audience page can read "In production" while the token server disagrees. Two readings that do not: the verification-status line on the Branding page, and google-health-mcp doctor, which fails loudly when the stored token records a short expiry.

3. Authorise

Put the downloaded client JSON where the server looks for it, unedited:

mkdir -p ~/.config/google-health-mcp
cp ~/Downloads/client_secret_*.json ~/.config/google-health-mcp/google_client.json
google-health-mcp auth

Your browser will warn that Google hasn't verified this app. That is expected, and the app is your own: these health scopes are classified restricted, and verification only matters above 100 users. Click Advanced, then Go to google-health-mcp (unsafe), and grant the scopes.

If you have more than one Google account, pick the one your health data is in. auth checks the account once the tokens are saved, and if Google reports it is not linked to Google Health, it says so and exits 1; run it again and choose the other account.

The flow listens on localhost:8081 for the callback, so that port must be free. It saves tokens to ~/.config/google-health-mcp/google_tokens.json, created 0600 on POSIX. Windows keeps only the owner-write bit, as its read-only attribute, and governs access by ACLs - so there the file is not restricted to your account, and what it grants is whatever its directory's ACLs pass down. Access tokens last an hour and refresh automatically. Refresh tokens do not rotate, so a token minted on a machine with a browser can be copied to a headless one.

If you authorised before publishing the app, re-run google-health-mcp auth afterwards: publishing does not extend a token already granted, and that one still expires after seven days.

4. Register with your MCP client

claude mcp add -s user google-health -- google-health-mcp

Running it with uvx instead: claude mcp add -s user google-health -- uvx google-health-mcp.

5. Check it

google-health-mcp doctor

Worth running before step 3 (Authorise) as well as after: it reports whether port 8081 can be bound and whether this host can open a browser, which are the two ways auth fails before it starts.

Offline and read-only: it reports which paths resolved where, whether the credential files are the right shape, whether the token is short-lived, whether the grant holds every permission this version reads, and whether the cache is being kept up to date.

doctor --json reports the same findings for a monitor to act on:

{
  "version": "1.6.0",
  "findings": [
    {
      "check": "stopped-series",
      "name": "hrv series",
      "severity": "warn",
      "detail": "No hrv since 2026-03-30, after rows on 28 of the 30 days before that.",
      "fix": "Re-sync that type alone (...)"
    }
  ],
  "counts": {"ok": 7, "warn": 1, "fail": 0}
}

The payload goes to stdout; logging goes to stderr, so a subprocess consumer should read the two separately.

Match on check, never on name or detail: the first is a stable identifier, the other two are prose and carry the data type. check is null for findings nothing consumes programmatically yet.

The exit code is 1 only when something is graded fail, in both formats - a warning never changes it, which is the reason this flag exists: a stopped data series is a warning, so the exit code alone cannot tell you about the one failure most worth watching for.

Check version before trusting an absent check. A release older than this one omits the field entirely and an older one still rejects --json and exits 2, so "no stopped-series finding" and "this build cannot report one" look identical without it. version is itself null when the package is run from a source tree with no installed distribution metadata - the payload is still emitted, since a diagnostic that dies on a half-configured install is worthless exactly when it is needed.

The payload names the resolved config, database and credential paths, the same way the text report does. That is deliberate - it is what makes a wrong-path setup diagnosable - but a consumer that forwards the payload off the machine is disclosing them. No credential values appear in either format.

6. First sync (optional)

Query tools sync on first use each day, so you can skip this. To pre-populate the cache, or to pull history older than it:

google-health-mcp sync --days 30
google-health-mcp sync --since 2023-10-01     # backfill

CLI usage

google-health-mcp                Start the MCP server (stdio transport)
google-health-mcp -V, --version  Print the installed package version
google-health-mcp auth           Interactive OAuth setup
google-health-mcp doctor         Check the setup and report what needs fixing
  --json                Emit the findings as JSON, for a monitor rather than
                        a person. Each finding carries a stable `check` name
                        to match on; the exit code is the same either way.
google-health-mcp sync           Sync data to the local cache
  --days N              Days of history for a first sync (default: 30)
  --types TYPE,...      Data types to sync (default: all). One or more of:
                        heart_rate, activity, exercises, sleep, weight, spo2,
                        hrv, azm, breathing_rate, skin_temperature,
                        core_temperature, cardio_fitness, food_log, ecg, irn,
                        account, height, exercise_routes
  --since YYYY-MM-DD    Fetch from this date, ignoring the incremental cursor
  --until YYYY-MM-DD    Inclusive end date for a --since window; together they
                        re-fetch exactly that window, to repair a gap in the
                        middle of the cache
google-health-mcp import         Import exported JSON data files
  --data-dir PATH       Directory containing the JSON files

MCP tool reference

Query tools sync on the first query of each day per data type, then read the cache.

All query tools except health_get_devices, health_get_lifetime_stats, health_get_profile and health_get_exercise_route accept:

• start_date - YYYY-MM-DD, YYYY-MM, or 30d (relative). Default: last 30 days, except health_get_height, whose default is the last ten years.
• end_date - YYYY-MM-DD. Default: today.
• live - if true, re-fetch this window from the API before reading the cache. A failed refresh is reported rather than silently answered from the cache.

health_get_profile takes only live.

health_get_exercises also takes exercise_type, a case-insensitive substring match on the workout name. Google names the workouts, so a value matching no workout name in your cache is refused with the cached names listed and the live=True hint, rather than answered as a period you did not train in. health_get_ecg also takes include_waveform: a trace is thousands of voltages, so the default response carries the classification, average rate, duration and a sample count instead.

health_get_exercise_route takes a workout's log_id. The route text comes back only with include_tcx, since it can run to hundreds of kilobytes. Only workouts recorded with GPS have a route, and a sync fetches routes for the workouts in its own window, so an older one needs health_sync with data_types="exercise_routes" and since. A route holds every position you recorded, start and end included: it is cached on disk like the rest, and reaches the model only when include_tcx is set.

If your authorisation lacks a permission this version reads, every successful tool response carries an authorisation note naming it, doctor reports it under the check missing-scopes, and sync prints it. The data types read under it are skipped rather than failed, so sync still exits 0. A grant does not gain permissions on refresh, so this is what an upgrade that adds one looks like until you run google-health-mcp auth again. doctor on the syncing host sees the new grant straight away; the tool note, and doctor on a cache-only host, catch up once the next sync has run. Run google-health-mcp sync after authorising to fetch the newly granted data at once; a query tool otherwise waits for the next day, since a skipped type counts as synced for that day.

health_sync

• data_types - all, or a comma-separated subset of the names listed under CLI usage above (irn is the irregular-rhythm notifications, and account is the profile, settings and irregular-rhythm enrolment). Default: all.
• days - days of history for a first sync (default: 30). Later syncs are incremental.
• since / until - fetch an exact window regardless of what is cached.

A sync adds and corrects, and never removes. An entry deleted in the app it was logged in simply stops appearing in the API, with nothing to say it was deleted, so a re-sync of that window leaves the cached row as it was. The API offers no tombstone, no change feed and no modified-since filter, so there is nothing to detect it by. In practice this reaches only the types you fill in by hand, since a device series is not deleted after the fact. Re-creating the cache is the way to clear one.

health_trends

• data_type - any cached type with a daily series; ECG readings and rhythm alerts are episodes and have no trend. Default: activity.
• period - weekly, monthly, quarterly. Default: monthly.
• start_date / end_date - default: the last 12 months.
• compare - two periods, e.g. last_30d vs previous_30d, 2026-03 vs 2026-02, 2026-Q1 vs 2025-Q4. When set, period, start_date and end_date are ignored.

OAuth scopes

Tick these read-only scopes on the Data Access page. All are under https://www.googleapis.com/auth/googlehealth.:

The console also offers reproductive_health.readonly, logged_symptoms.readonly and mindfulness.readonly, which this package does not request. As of September 2026 the cycle, ovulation-test, symptom and mood data types can be written but not read, and the API has no mindfulness data type, so there is nothing to read under them.

Read the list off the console, not off the published scope page - read-only scopes exist that appear in neither Google's documentation nor the API's own discovery document, and the discovery document omits nutrition.readonly outright. To request fewer, tick fewer on the Data Access page and remove them from GOOGLE_SCOPE_READERS in config.py before authorising, which needs a source checkout rather than a pip or uvx install; the data types under a scope you leave out are skipped by every sync. A grant does not gain scopes on refresh, so widening the list later means running auth again.

Configuration

Offline / cache-only mode

By default the server syncs on demand, so no cron job is needed. Set GOOGLE_HEALTH_MCP_OFFLINE=1 to run as a pure reader instead:

• No credentials are required - the server never opens the token file.
• No network call is made. Auto-sync is off, and live=True, health_get_devices and health_sync return a clear "offline mode" message rather than reaching the API.
• Query tools serve the cache, tagged "offline_mode": true.

Typical uses:

• Several machines, one cache - one host runs google-health-mcp sync from cron or systemd against a shared database; the others set GOOGLE_HEALTH_MCP_OFFLINE=1, point GOOGLE_HEALTH_MCP_DB_PATH at the same file, and only read.
• CI and privacy - run queries with no network access and no credentials.

Rate limits

Google applies a per-user request quota, documented at developers.google.com/health/rate-limits. Ordinary syncing is nowhere near it: a day's update is a handful of requests, and a measured three-year backfill of every data type was around 250. If a sync is cut short, that data type is recorded as a partial sync and the next run resumes from its cursor rather than starting over.

Querying from the cache - the default - costs no quota at all.

Data safety

Your health data stays on your machine: this server has no backend, sends nothing anywhere, and talks only to Google's API with your own credentials.

The repository ships a pre-commit hook that refuses to commit database files, anything under config/, and large files; CONTRIBUTING.md says how to install it.

Importing existing data

If you already have health data as JSON files, from an export or a script of your own:

google-health-mcp import --data-dir /path/to/json/files/

Expected file names: heart_rate.json, activity.json, exercises.json, sleep.json, weight.json, spo2.json, hrv.json. See src/google_health_mcp/importer.py for the shape each one expects. Import covers those seven types; everything else arrives by sync.

Contributing

See CONTRIBUTING.md for development setup, the test workflow, and the pre-commit hook. Changes are tracked in CHANGELOG.md.

License

GPL-3.0-or-later