google-health-mcp-aws

An MCP (Model Context Protocol) server that gives Claude access to sleep and recovery data from a Fitbit or Pixel Watch, through the Google Health API (the successor of the Fitbit Web API). It runs on AWS Lambda as a claude.ai custom connector, so it works in claude.ai on the web, on mobile and in Claude Desktop. Several people can share one deployment, and each signs in with their own Google account.

It's meant for a training coach in Claude: HRV, resting and sleeping heart rate, SpO2, breathing rate, skin temperature and sleep, compared with your own baseline.

• "How recovered am I this morning?"

• "How has my HRV trended over the last month?"

• "Show me last night's sleep in detail. When was my heart rate lowest?"

• "Did my sleep get worse during the training block that started on the 1st?"

Tools

All tools are read-only. Nightly metrics belong to the date you woke up, as in the Fitbit app.

Related MCP server: whoop-mcp

Setup

You need an AWS account, a Google account, and a Fitbit or Pixel Watch whose data is in that Google account. A Fitbit account has to be moved to a Google account for the Google Health API to see it.

Install Node.js 20+, the AWS CLI and the SAM CLI, then sign in to AWS with a default region:

winget install OpenJS.NodeJS.LTS
winget install Amazon.AWSCLI
winget install Amazon.SAM-CLI
aws configure        # or: aws login
npm install

1. Create the Google Cloud project

1. In the Google Cloud console, create a project, for example google-health-mcp.

2. Enable the Google Health API in that project.

3. Open Google Auth Platform and click Get started. Give the app a name (people see it on Google's consent screen), set a support email, and choose External as the audience.

4. Under Data access → Add or remove scopes, add these four scopes:

   • openid

   • .../auth/userinfo.email

   • .../auth/googlehealth.sleep.readonly

   • .../auth/googlehealth.health_metrics_and_measurements.readonly

5. Under Audience, click Publish app to set it In production. Google doesn't need to verify it for up to 100 users.

2. Deploy

npm run deploy

Confirm the changeset when asked. The output ends with two values:

• McpServerUrl, which people add to claude.ai.

• GoogleRedirectUri, which the Google OAuth client needs next.

sam list stack-outputs shows them again later.

3. Create the Google OAuth client

1. In Google Auth Platform, go to Clients → Create client and choose application type Web application.

2. Under Authorized redirect URIs, add the GoogleRedirectUri from the deploy output. It ends in /oauth/callback.

3. Create the client and download its JSON file.

4. Store the client in SSM, then delete the downloaded file:

   npm run admin -- google-client path/to/client_secret_....json

5. Allow your own Google account to connect:

   npm run admin -- allow you@gmail.com

4. Connect Claude

In claude.ai, go to Settings → Connectors → Add custom connector and paste the McpServerUrl (it ends in /mcp). Click Connect and sign in with Google. On the consent screen, allow both sleep data and health metrics.

The connector then works on web and mobile, and in Claude Desktop. To use it in a training-coach project, enable it in that project's chat.

Adding someone

1. Run npm run admin -- allow their@gmail.com with the Google account that holds their Fitbit data.

2. Send them the McpServerUrl. They add the connector in their own claude.ai account and sign in with Google. On the Free plan, claude.ai allows one custom connector.

Each person sees only their own data. Their Google tokens are stored nowhere on the server; they travel inside the encrypted tokens Claude holds for their connection.

Day to day

Lambda reads its settings once per cold start, so every admin change also replaces the function's running instances. That way changes take effect immediately.

People can also remove the server's access themselves, under Google Account → Third-party connections. Claude then asks them to connect again.

How it works

• Sign-in: claude.ai custom connectors only support OAuth, so the server is a small OAuth server that passes sign-in on to Google. When someone connects, the server sends their browser to Google's consent screen. It then checks their Google email against the allowlist and hands Claude a code.

• Tokens: Claude's access token carries the person's Google access token, and its refresh token carries their Google refresh token, all encrypted (AES-256-GCM) with a key only the server has. Every hour Claude refreshes its token, and the server gets a new Google access token. If Google no longer accepts the grant, Claude asks the person to reconnect.

• No database: nothing is stored per person. The signing secret, the allowlist and the Google OAuth client are SecureString parameters in SSM Parameter Store, so they never appear in the function's configuration or in the repo.

• Data: tools use the Google Health API's reconcile endpoint, which merges data from all sources (Fitbit, Pixel Watch, Health Connect apps) so a night isn't counted twice.

• Cost: normally $0, within Lambda's always-free allowance (1M requests and 400,000 GB-seconds a month). Set up an AWS budget alert anyway.

Running locally

For testing, or for MCP clients on your own machine, the same server runs locally with settings from .env:

1. Add http://localhost:3000/oauth/callback to the Google OAuth client's authorized redirect URIs.

2. Copy .env.example to .env and fill it in.

3. Run:

   npm run build
   npm start

The server listens on http://localhost:3000/mcp. For example: claude mcp add --transport http google-health http://localhost:3000/mcp.

Known limitations

• Data freshness: data only reaches Google after the watch syncs with the Fitbit app, and some metrics (HRV, skin temperature) are computed a while after waking. This morning's values can be missing at first.

• Reconnecting: Google ends grants after 7 days while the app is in Testing, and after 6 months without use. Some integrations also report grants ending unexpectedly (home-assistant/core#182964). In all these cases Claude asks you to connect again.

• Unverified app: see the note in step 1. The limit is 100 people.

License

MIT