Google Health

Google Health — read-only data extractor

Read your Google Health data from any agent. This is a generic, read-only client for the Google Health v4 API: it authenticates, pulls data points for any data type, and prints JSON. It does no filtering, merging, or interpretation — you get the data and parse whatever you care about. Single static binary — no Python or other runtime.

⚠️ Read-only. It requests only read-only Google Health scopes and never writes your health data.

🔒 Privacy — sensitive health data leaves on stdout. Every data list, rollup daily, sessions, and api get prints your personal health information (heart rate, sleep, weight, exercise, blood oxygen, profile, …) to stdout as JSON. In an agent or pipeline, stdout may be logged, summarized, persisted, or transmitted to other tools, model providers, or third parties. Treat this output as sensitive PII: only run it in contexts you trust to handle health data, and do not persist or forward it unless you intend to. The tool never sends your data anywhere itself (it only talks to Google's API over HTTPS) — what happens to stdout is the caller's responsibility. Every data-emitting command also prints a one-line privacy notice to stderr at run time as an execution-time reminder (it stays on stderr, never on the JSON stdout).

🧭 Data minimization & consent. Request only the data types and the narrowest time window you actually need (--days/--date/--from/--to, not --all); prefer the typed data/rollup/sessions commands over api get; and don't persist or forward the output beyond the task at hand. Run this only against an account whose owner has knowingly consented to having their health data read by this agent — the one-time auth login is the owner consenting to read-only access for the tool, not consent for a downstream agent to collect, retain, or transmit that data more broadly.

🩺 doctor prints local environment metadata. The doctor JSON includes your token-cache and config file paths, the configured account (user), and the API base URL — local filesystem layout and account identifiers rather than health data, but still local environment metadata that an agent or pipeline may log, forward, or persist. It exists to diagnose your own setup; treat its output like any other local path/account detail and don't paste it into untrusted contexts.

Setup (one time)

Unlike a simple email/password tool, Google Health uses OAuth, so first-time setup needs a human in a browser once. After that it runs headless (the token auto-refreshes).

1. Install the binary:

   # A) Download a release for your OS/arch and put it on PATH:
   #    https://github.com/stozo04/google-health-cli/releases
   # B) Or with Go (1.25+):
   go install github.com/stozo04/google-health-cli/cmd/google-health-cli@latest
   

2. Create a Google Cloud OAuth client (Desktop app, read-only Health scopes, ~10 min, no fees). Full walkthrough: the repo's OAUTH_SETUP.md.
3. Provide the client via GOOGLE_HEALTH_CLIENT_ID / GOOGLE_HEALTH_CLIENT_SECRET (or a config.json with those keys in the working directory).
4. Authorize once (opens a browser, asks a human to consent to read-only access):

   google-health-cli auth login
   google-health-cli doctor        # expect "tokenValid": true
   

   The token is cached 0600 and auto-refreshed; every later run is headless.

Credentials

config.json (gitignored) is an alternative to the env vars:

{
  "client_id": "PASTE_CLIENT_ID",
  "client_secret": "PASTE_CLIENT_SECRET"
}

🔑 Protect your credentials. The OAuth client_id/client_secret and the cached token are sensitive secrets stored on disk in plaintext (config.json / token.json, written 0600). They're gitignored — keep it that way: never commit them, copy them into a shared folder, back them up to an untrusted location, or paste them into logs or chat. Rotate the OAuth client (and run auth logout) if one may have leaked.

Commands

Discover the data types

google-health-cli types list                 # all types ( * = supports list )
google-health-cli types describe heart-rate   # one type's record shape, scope, time field

Read data points (the core)

google-health-cli data list heart-rate --days 1        # last 1 day
google-health-cli data list sleep --days 14             # last 14 days
google-health-cli data list daily-resting-heart-rate --days 30
google-health-cli data list steps --date 2026-06-16     # one civil day
google-health-cli data list weight --from 2026-01-01T00:00:00Z --to 2026-07-01T00:00:00Z
google-health-cli data list distance --all              # everything, no time filter

stdout is always a JSON array of the raw data points (verbatim from the API); a one-line count goes to stderr. Window flags (precedence: --all > --from/--to > --date/--days): --date today|yesterday|YYYY-MM-DD, --days N (default 7), --from/--to (RFC3339, together), --all (no filter).

Example (data list daily-resting-heart-rate):

[
  {
    "dataSource": { "platform": "FITBIT", "device": { "displayName": "Google Pixel Watch 4 (45mm)" } },
    "dailyRestingHeartRate": { "date": { "year": 2026, "month": 6, "day": 17 }, "beatsPerMinute": "78" }
  }
]

A few types are roll-up/reconcile-only and can't be listed (types list marks listable ones with *); read those with rollup daily instead (below). One type — daily-heart-rate-zones — supports neither list nor rollup daily; it is reachable only via the api get escape hatch.

Daily roll-ups (server-side totals)

google-health-cli rollup daily steps --days 7           # one reconciled total per civil day
google-health-cli rollup daily active-minutes --days 7   # a roll-up-only type (no `data list`)

rollup daily returns the API's daily roll-up rows — one value per civil (local) calendar day — instead of the raw per-minute/per-sample points data list returns. It's the cheap way to get daily totals (e.g. a steps total per day rather than a per-minute dump) and the only way to read the roll-up-only types that have no list op (active-minutes, total-calories, floors, calories-in-heart-rate-zone, time-in-heart-rate-zone). stdout is a JSON array of the raw roll-up rows; an N daily rollup(s) count goes to stderr. Each value is reconciled across data sources (and excludes off-wrist intervals), so it is not a naive sum of data list points. Window flags mirror data list minus --all (the range is required).

[
  {
    "civilStartTime": { "date": { "year": 2026, "month": 6, "day": 16 }, "time": {} },
    "civilEndTime":   { "date": { "year": 2026, "month": 6, "day": 17 }, "time": {} },
    "steps": { "countSum": "8000" }
  }
]

Parsed exercise sessions (convenience)

google-health-cli sessions --days 14 --json    # flattened exercise rows
google-health-cli sessions --raw               # raw exercise data points

[
  { "date": "2026-06-16", "exercise_type": "ELLIPTICAL", "duration_min": 35,
    "avg_hr": 130, "calories": 236, "platform": "FITBIT" }
]

sessions returns all exercise types (no filtering); the caller keeps what it wants.

Raw read-only API access

google-health-cli api get /v4/users/me/profile     # GET any read-only v4 path
google-health-cli api get /v4/users/me/settings

⚠️ Sensitive endpoints — broader than the typed metrics. api get can read any read-only v4 path, including users/me/profile and users/me/settings — account, identity, and settings data that the typed commands don't surface. Its responses are personal data printed to stdout, so the same Privacy caution above applies, and the broad reach makes over-collection easy. It stays GET-only (no path can write or delete) and is constrained to the read-only v4 surface (/v4/...): a non-v4/ path, an absolute URL, or a .. traversal — literal or percent-encoded (%2e%2e, ..%2f) — is rejected (exit 64) before any request is made. Still, prefer the typed data/rollup/sessions commands and reach for api get only when you need an endpoint they don't model.

Full command reference

Conventions

• stdout is parseable; human hints, counts, and logs go to stderr, never interleaved.
• Exit codes: 0 success, 2 auth/API failure (incl. doctor when unauthenticated), 64 usage error (incl. an expired token with no OAuth client credentials to refresh it), 78 config error. An expired token plus an undiscoverable/credential-less config fails fast with an actionable message — not the cryptic oauth2 "Could not determine client ID" error.
• Read-only: six read-only googlehealth.*.readonly scopes; no write operations exist.
• Secrets: config.json and the token cache are gitignored — never commit them.
• Time-filter formats are handled for you per data type (civil wall-clock, RFC3339 instant, or date-only); a --date/--days window always covers your local calendar days. If a type rejects server-side filtering, re-run with --all.

See the repo's AGENTS.md for the exact --json shapes and exit-code contract.