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
| Tool | Data |
|---|---|
health_get_heart_rate | Resting heart rate |
health_get_activity | Steps, calories, distance, floors |
health_get_exercises | Workouts (name, duration, heart rate, calories) |
health_get_sleep | Duration, stages, sleep period |
health_get_weight | Weight, body fat % |
health_get_spo2 | Nightly blood oxygen saturation |
health_get_hrv | Heart rate variability (RMSSD) |
health_get_azm | Active zone minutes, with the per-zone breakdown |
health_get_breathing_rate | Nightly breaths per minute |
health_get_skin_temperature | Nightly variation from your baseline, and the absolutes behind it |
health_get_core_temperature | Body temperature readings you logged by hand |
health_get_cardio_fitness | VO2 max, where the device reports it |
health_get_food_log | Food calories and water, where logged |
health_get_ecg | Electrocardiograms: classification, average rate, duration, waveform on request |
health_get_irregular_rhythm | Irregular-rhythm notifications and the windows that triggered them |
health_get_devices | Paired devices, battery level, last sync |
health_get_lifetime_stats | Totals and best days over the cached history, with its coverage |
health_trends | Aggregated averages and period comparisons |
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
Or run it without installing, in which case every google-health-mcp ... command you run below becomes uvx google-health-mcp ...:
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.
- Project. Create a project at console.cloud.google.com/projectcreate and select it.
- API. Enable Google Health API on the API Enablement page.
- 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.
- Audience. Under Test users, add your own Google account. Skipping this fails sign-in with
403: access_denied. - Data Access. Click Add or remove scopes, search for "Google Health API", and tick the read-only scopes listed under OAuth scopes below.
- 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.
- 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:
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.
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
Running it with uvx instead: claude mcp add -s user google-health -- uvx google-health-mcp.
5. Check it
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, and whether the cache is being kept up to date.
doctor --json reports the same findings for a monitor to act on:
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:
CLI usage
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 and health_get_lifetime_stats, which take no arguments, accept:
start_date-YYYY-MM-DD,YYYY-MM, or30d(relative). Default: last 30 days.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_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_sync
data_types-all, or a comma-separated subset of the names listed under CLI usage above (irnis the irregular-rhythm notifications). 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_dateandend_dateare ignored.
OAuth scopes
Tick these read-only scopes on the Data Access page. All are under https://www.googleapis.com/auth/googlehealth.:
| Scope | Data accessed |
|---|---|
activity_and_fitness.readonly | Steps, distance, floors, calories, workouts, active zone minutes |
health_metrics_and_measurements.readonly | Heart rate, HRV, SpO2, breathing rate, weight, body fat, temperature, VO2 max |
sleep.readonly | Sleep sessions and stages |
nutrition.readonly | Food and water logs |
ecg.readonly | Electrocardiograms |
irn.readonly | Irregular-rhythm notifications |
settings.readonly | Paired devices |
location.readonly and profile.readonly are two the console offers that this package deliberately does not request, because nothing here reads either - the first is the GPS track recorded during an exercise.
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 edit GOOGLE_SCOPES in config.py before authorising, which needs a source checkout rather than a pip or uvx install. A grant does not gain scopes on refresh, so widening the list later means running auth again.
Configuration
| Variable | Default | Description |
|---|---|---|
GOOGLE_HEALTH_MCP_CONFIG_DIR | ~/.config/google-health-mcp/ | Directory holding the OAuth client and tokens |
GOOGLE_HEALTH_MCP_DB_PATH | ~/.local/share/google-health-mcp/google_health.db | SQLite cache |
GOOGLE_HEALTH_MCP_OFFLINE | unset | If truthy (1, true, yes, on), run as a cache-only reader |
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_devicesandhealth_syncreturn 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 syncfrom cron or systemd against a shared database; the others setGOOGLE_HEALTH_MCP_OFFLINE=1, pointGOOGLE_HEALTH_MCP_DB_PATHat 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:
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.