Your WHOOP recovery, sleep, strain and workouts in Claude, ChatGPT or other MCP apps. Self-hosted.
This MCP server provides access to WHOOP recovery, sleep, strain, and workouts within Claude, ChatGPT, or other MCP apps. It is designed to be self-hosted, enabling developers to integrate WHOOP-related data flows into their existing MCP environment.
🛠️ Key Features
WHOOP data access for recovery, sleep, strain, and workouts
Works with Claude, ChatGPT, and other MCP applications
Self-hosted deployment
🚀 Use Cases
Display WHOOP recovery and sleep information in an MCP chat or workflow
Include WHOOP strain metrics in developer-built assistants
Retrieve WHOOP workouts data for assistant-driven summaries or automation
⚡ Developer Benefits
Integrates WHOOP recovery, sleep, strain, and workouts into MCP-based apps
Supports use across multiple MCP clients (e.g., Claude, ChatGPT)
⚠️ Limitations
Details such as tool count or supported operations are not provided in the available data
Your WHOOP data in whichever AI you use, or, soon, in your own code. Self-hosted, private, and open source.
Ask Claude, ChatGPT or another MCP app "How did I sleep this week?" and get the answer from your own recovery, sleep, strain and workouts. You run the server yourself: it fetches your data from WHOOP when you ask and keeps no copy, and only the apps you've allowed can reach it.
This is an independent open-source project. It uses the WHOOP API to access data from WHOOP products, and is not affiliated with, endorsed by, or sponsored by WHOOP.
Two ways in
Talk to your data. Deploy the server, connect it to your AI, and ask. Start with Setup, then Add to your AI, which also lists the apps tested so far.
Build with it. The WHOOP client inside this server (typed, tested, and careful with WHOOP's single-use refresh tokens) is being split out as a library for your own code. It isn't published yet: watch this repository's releases to hear when it is.
Workouts: activity, local start time, duration, strain, heart rate, calories, and time in heart-rate zones 4–5
Live data: every answer is fetched from Whoop when you ask, so it's always current. The server stores only its sign-ins and your encrypted Whoop tokens, never your health data
Private by default: each app signs in with a password you choose (OAuth 2.1), and only after you tick a box allowing it, so nobody else can read your data
Your choice of AI: tested live with Claude and ChatGPT; other apps that sign in with OAuth should work the same way (see compatibility)
MCP Tools
Tool
Description
get_today
Morning briefing with recovery, sleep, and strain
get_recovery_trends
Recovery patterns over time with HRV/RHR
get_sleep_analysis
Sleep trends: time asleep, performance, and efficiency
get_strain_history
Daily strain and calorie trends
get_workouts
Recent workouts with activity, duration, strain, heart rate, and calories
get_auth_url
Link to connect your Whoop account (works once, expires in 10 minutes)
Privacy Policy: this project's PRIVACY.md, https://github.com/yuridivonis/whoop-mcp-server/blob/main/PRIVACY.md, or your own adapted copy if you run the server for someone else. People see this link when they approve the app.
Redirect URL: your server's callback, e.g. https://your-app.up.railway.app/callback. If you don't have the address yet, fill this in after step 2 creates it.
Scopes: read:recovery, read:cycles, read:sleep, and read:workout. The server doesn't use the others. The login also asks for offline, which lets the server renew its Whoop access without you logging in again; the dashboard doesn't list it.
Webhooks: leave empty.
Note your Client ID and Client Secret.
2. Deploy
Every release is published as a ready-made image, ghcr.io/yuridivonis/whoop-mcp-server. Deploy that: there's no need to fork this repository unless you want to change the code (see Changing the code). The steps below use Railway; to run it anywhere else, see Docker.
In a Railway project, click New, choose Docker Image, and enter ghcr.io/yuridivonis/whoop-mcp-server:1.4.1. Then, in the service's Settings → Networking → Public Networking, choose Generate Domain: that's your server's address, your-app.up.railway.app below.
MCP_AUTH_PASSWORD: the password each AI app asks for when you connect it. Generate one with openssl rand -base64 24 and keep it in your password manager. The server refuses to start without it (at least 16 characters).
ENCRYPTION_SECRET (optional, recommended): generate one with openssl rand -base64 32. It encrypts your stored Whoop tokens, so rotating the Whoop client secret later won't disconnect your account.
Add a volume mounted at /data. It holds the sign-ins and your encrypted Whoop tokens; without it, every redeploy signs your apps out and disconnects Whoop.
Turn on updates: in the service's Settings, under Source, choose Configure Auto Updates, pick minor updates and patches, and a maintenance window (for example Night). Railway then moves the service to each new 1.x release by itself, and on the Pro plan backs up the volume first.
Deploy, then open https://your-app.up.railway.app/health to check it's running.
Already running a fork on Railway? Switch it to the image: open the service's Settings, change Service Source to ghcr.io/yuridivonis/whoop-mcp-server:1.4.1, and turn on auto updates as in step 4. Keep the same variables and volume, so Whoop stays connected. If your fork is older than 1.3.0, read Upgrading to 1.3.0 first: every app signs in once more. Your fork is then no longer used.
3. Connect your AI
Add your server's address, https://your-app.up.railway.app/mcp, to your AI app as a custom connector. It opens your server's sign-in page: check it names the app you're connecting, enter your MCP_AUTH_PASSWORD, and tick the box allowing it to read your Whoop data. For example, in Claude.ai: Customize → Connectors → + → Add custom connector.
Add to your AI has the steps for Claude, ChatGPT, Claude Code, Cursor, VS Code and Windsurf, and which have been tested. Apps stay signed in across redeploys; anyone without the password gets 401 Unauthorized from /mcp.
4. Connect your Whoop account
In a chat, ask your AI to connect Whoop. It calls get_auth_url and gives you a link.
Open the link, log in to Whoop, and authorize the app. You're redirected back, and your AI can answer right away.
Ask away: "How did I sleep last night?"
Upgrading to 1.3.0
1.3.0 stops keeping a copy of your Whoop data, and asks you before each app receives it. To upgrade:
Get the new version: with the image, Railway's auto updates do it for you, or change the tag under Service Source to the new version (Docker with :1: pull it again and restart). With a fork, use GitHub's Sync fork button, then redeploy, or switch to the image as described in Deploy.
On its first start, the server deletes the recovery, sleep, strain and workout data earlier versions stored, and rewrites the database file so none of it is left on disk. Your Whoop connection is kept. The log says Deleted the WHOOP data stored by an earlier version.
Every app is signed out once. The next time you use one, it opens the sign-in page: enter your password and tick the box allowing it to read your Whoop data.
sync_data is gone. If your app still lists it, remove the connector and add it again.
If your host keeps backups or snapshots of the volume, delete the ones from before the upgrade. They still hold the old copy.
There's no going back to 1.2.x on the upgraded database: it can't sign apps in with the new sign-in tables.
Upgrading from 1.0.0
1.1.0 puts a sign-in in front of /mcp. Version 1.0.0 had no authentication there, so any 1.0.0 server that worked with Claude over HTTP served its data to anyone who knew the URL. (Unmodified 1.0.0 also had a request-parsing bug that stopped Claude from connecting over HTTP at all; 1.1.0 fixes both.) To upgrade:
Update your fork (GitHub's Sync fork button, or merge the upstream main branch), or switch the service to the image as described in Deploy.
Set MCP_AUTH_PASSWORD in your Railway variables (see Setup, step 2). Without it, the new version won't start. That's deliberate.
Redeploy.
In Claude.ai → Settings → Connectors, remove the Whoop connector and add it again with the same URL. Claude shows the sign-in page once.
If a tool says your Whoop authorization expired, run get_auth_url once to reconnect.
Optional: in your Whoop app, untick read:profile and read:body_measurement. 1.1.0 no longer uses them.
If your 1.0.0 server worked with Claude on a public URL, assume your data could have been read. As a precaution, rotate your client secret in the Whoop developer dashboard, update WHOOP_CLIENT_SECRET, and run get_auth_url once afterwards. Unless ENCRYPTION_SECRET is set, the stored Whoop tokens were encrypted with the old client secret. The server starts anyway and treats Whoop as disconnected until you reconnect.
Security
/mcp only answers signed-in clients. Sign-in codes and refresh tokens work once and are stored as hashes; if one is ever used twice, the whole sign-in is revoked.
Failed sign-ins are limited to 10 per address every 15 minutes, and 50 per hour in total.
Sign-in codes only go to Claude, ChatGPT, desktop apps on your own computer, or web clients you add with MCP_ALLOWED_REDIRECT_HOSTS. The sign-in page shows where you'll return: only sign in if you started the connection yourself.
Each app is signed in only after you tick a box allowing it to read your Whoop data.
Whoop tokens are encrypted at rest (AES-256-GCM). Your health data isn't stored: the server fetches it from Whoop for each question and sends it only to the client you signed in.
Changing MCP_AUTH_PASSWORD signs every client out.
The server asks Whoop only for recovery, cycles, sleep, and workouts.
See SECURITY.md for the full security model and how to report a vulnerability privately, and PRIVACY.md for what a deployment stores and shares.
WHOOP's terms: what the server does, and what's up to you
When you deploy this server, you register your own Whoop developer app, so you are the developer (the "Company") under WHOOP's API Terms of Use, effective 6 October 2026. The terms number their sections but not the paragraphs inside them, so this cites each paragraph by its section and heading. It's a summary of the obligations that touch how the server handles data, not legal advice, and not a promise that any deployment complies: read the terms before you deploy.
WHOOP's terms
What they require
What the server does, and what's up to you
4. WHOOP Data, Prohibitions on WHOOP Data
Explicit opt-in consent before WHOOP data reaches a third party
Every app you connect is a third party. The sign-in page names where the data goes, and doesn't sign the app in until you tick the box allowing it.
4. WHOOP Data, Prohibitions on WHOOP Data
No databases or permanent copies, and no cached copies kept longer than WHOOP's cache headers allow
Nothing is stored. Each answer is fetched from WHOOP when you ask; tools that ask at the same moment share one request, and nothing is kept once it's done.
4. WHOOP Data, Prohibitions on WHOOP Data
No using WHOOP data to create, develop, test, train, fine-tune or improve AI
The server trains nothing, and this project's tests use synthetic data only. Before you connect an app, turn off any setting that lets its provider use your conversations to improve its models.
2. Company Applications, Application Security
WHOOP data encrypted in transit and at rest; security incidents reported to WHOOP within 48 hours
The server refuses to run on a public address without https, and the only WHOOP data it stores is your encrypted tokens. Reporting an incident is your job: see below.
1. Use of WHOOP APIs, Permitted Access
One set of WHOOP credentials per application
Give each deployment its own Whoop developer app.
3. Restrictions; Confidentiality, Confidentiality
Developer credentials kept confidential, and never embedded in open-source projects
The server reads them from environment variables. Never commit them to a repository, including a fork.
3. Restrictions; Confidentiality, API Prohibitions
No medical, legal or other professional advice, and no medical devices
The tools report WHOOP's numbers. They don't give advice or diagnose anything, and neither should anything you build on them.
If you run it for someone else
Each deployment serves one Whoop account. If you deploy it for someone else, they're your end user, and under 2. Company Applications you're responsible for:
Consent: they connect their own Whoop account through Whoop's login, and tick the box for each app themselves, so they need the server password (MCP_AUTH_PASSWORD). Don't tick it for them (End User Authorization and Consent).
A privacy policy: adapt PRIVACY.md with your contact details, and link it from your Whoop app (End User Privacy).
Support: give them an easy way to reach you (End User Authorization and Consent, Application Support).
Security incidents: if anyone gets unauthorized access to their data, notify WHOOP within 48 hours of discovering it, at security-notifications@whoop.com, and tell them as the law requires (Application Security). See SECURITY.md.
Updates: keep the deployment on a supported version.
AI: never use their data to test or improve any AI system, including your own experiments.
Docker
Each release is published as an image for amd64 and arm64 on GitHub's container registry. It runs the same server as the Railway setup above:
On a server with a public https address: set WHOOP_REDIRECT_URI to that address's /callback, and connect your AI app to its /mcp, as with Railway.
On your own computer: Whoop's login still needs an https address, so point a tunnel at port 3000 (see below) and use the tunnel's /callback. Add -e PUBLIC_URL=http://localhost:3000, so MCP clients on the same computer connect to http://localhost:3000/mcp.
The sign-ins and Whoop tokens live in the whoop-data volume, so restarts and upgrades keep you connected.
Tags::1 always points to the newest 1.x release, so pulling it again (or redeploying) picks up fixes and new features without breaking changes. :1.3.0 and the like pin one exact version; :latest follows every release, including a future 2.0.
To check that an image was built by this repository's release workflow, run gh attestation verify oci://ghcr.io/yuridivonis/whoop-mcp-server:1 --owner yuridivonis.
The server is also listed in the official MCP Registry as io.github.yuridivonis/whoop-mcp-server.
Running on Your Own Computer
Requires Node.js 22 or later.
bash
# Install dependencies
npm install
# Create .env file (npm run dev loads it)cat > .env << EOF
WHOOP_CLIENT_ID=your_client_id
WHOOP_CLIENT_SECRET=your_client_secret
# Whoop needs an https address: use your tunnel's (see below)
WHOOP_REDIRECT_URI=https://your-tunnel.example.com/callback
MCP_AUTH_PASSWORD=choose-a-local-password
MCP_MODE=http
EOF# Run in development mode (restarts on changes)
npm run dev
# Run the tests and the type check
npm test
npm run typecheck
Whoop's redirect URLs must be https (or an app scheme), so a server on your computer needs an https tunnel, for example cloudflared tunnel --url http://localhost:3000 or ngrok http 3000. Then:
Set WHOOP_REDIRECT_URI to the tunnel's /callback address.
Add that address to your Whoop app.
Connect your MCP client to the tunnel's /mcp address.
Quick tunnels get a new address every time they start, so you'd repeat steps 1 to 3; a named tunnel keeps one address. The tunnel provider carries the traffic, including the tools' answers.
MCP_MODE=stdio runs the server for MCP clients that start it as a local command. It has no sign-in, because only the app that started it can reach it. It can't receive the Whoop login either, so connect Whoop once with the server in http mode and the same DB_PATH, stop it, then start the stdio server. Don't run both at once: Whoop replaces the refresh token on every use, so two servers sharing one database log each other out.
Changing the code
To run your own changes, fork this repository and deploy the fork instead of the image: on Railway, New → GitHub Repo, which builds the Dockerfile. A fork doesn't update itself: to pick up new releases, use GitHub's Sync fork button and redeploy, and merge your changes as you go. If you don't need changes, the image is simpler and keeps itself up to date.
Environment Variables
Variable
Description
Default
WHOOP_CLIENT_ID
Whoop OAuth client ID
Required
WHOOP_CLIENT_SECRET
Whoop OAuth client secret
Required
WHOOP_REDIRECT_URI
OAuth callback URL
http://localhost:3000/callback
MCP_AUTH_PASSWORD
Password for the sign-in page that protects /mcp (16+ characters)
Required in http mode
PUBLIC_URL
Public address of the server, if it differs from WHOOP_REDIRECT_URI's. AI apps must connect to PUBLIC_URL/mcp.
Origin of WHOOP_REDIRECT_URI
ENCRYPTION_SECRET
Key for encrypting stored Whoop tokens
WHOOP_CLIENT_SECRET
MCP_ALLOWED_REDIRECT_HOSTS
Extra web clients allowed to receive sign-in codes, as host names separated by commas (e.g. app.example.com). Claude, ChatGPT, and desktop apps on your own computer (local addresses, and Cursor, VS Code, and Windsurf links) are always allowed.
None
TRUST_PROXY
Proxies allowed to report the client's IP (used by the sign-in rate limits): a hop count, false, or addresses/subnets
1 on Railway, otherwise false
DB_PATH
SQLite database path (sign-ins and encrypted Whoop tokens)
./whoop.db
PORT
HTTP server port
3000
MCP_MODE
http for a server, or stdio for an MCP client that starts it as a local command (see Running on Your Own Computer)
http
UPDATE_CHECK
Once a day, ask GitHub for the latest release number, and end get_today's answer with a one-line notice when a newer version is out. The request carries nothing about you or your data. false turns it off.
true
Architecture
code
┌─────────────────────────────────────────────────┐
│ Your AI app (Claude, ChatGPT, ...) │
│ "How did I sleep last night?" │
└────────────────────────┬────────────────────────┘
│ signs in once (OAuth 2.1),
│ then calls tools on /mcp
▼
┌─────────────────────────────────────────────────┐
│ Whoop MCP Server │
│ │
│ ┌─────────────┐ ┌──────────────────┐ │
│ │ Sign-in │─────►│ SQLite Database │ │
│ │ (OAuth 2.1) │ │ - sign-ins │ │
│ └─────────────┘ │ - Whoop tokens │ │
│ ┌─────────────┐ │ (encrypted) │ │
│ │ MCP tools │ └──────────────────┘ │
│ └──────┬──────┘ ▲ │
│ ▼ │ │
│ ┌─────────────┐ │ │
│ │ Whoop API │─── tokens ────┘ │
│ │ Client │ (no health data is stored) │
│ └─────────────┘ │
└─────────┬───────────────────────────────────────┘
│ Whoop OAuth + API v2, live on every call
▼
┌─────────────────────────────────────────────────┐
│ Whoop API │
└─────────────────────────────────────────────────┘
Whoop API Endpoints Used
GET /v2/cycle - Physiological cycles (strain data)
GET /v2/recovery - Recovery scores
GET /v2/activity/sleep - Sleep records
GET /v2/activity/workout - Workout records
Contributing
Issues and pull requests are welcome. Before opening a pull request, run npm test and npm run typecheck; CI runs both, along with a Docker smoke test.
Synthetic data only. The tests run against a fake Whoop API that serves made-up records (test/fake-whoop.ts), and future evaluations will too. Never put real Whoop data, yours or anyone else's, in tests, fixtures, issues or pull requests: WHOOP's terms forbid using it to test AI systems, and it's personal health data.
