Totem README

Totem gives Claude (or any MCP-compatible AI) full read + write access to your wearable data. Today it speaks Whoop — every metric, through Whoop's private iOS API.

_{▶ 2-min demo — the full totem cloud flow: install → Whoop login → Fly deploy → Claude connector → first query.}

55 tools, structured zod-validated outputs, bundled catalogs (372 exercises, 308 behaviors, 203 sports, 311 endpoints), write-safety harness, automatic Cognito token refresh, bounded in-memory read cache. TypeScript 6, Node 24, 278 tests.

Totem is becoming a universal wearables bridge. Whoop is the first adapter — every metric, fully wired. Fitbit, Apple Watch, and Garmin are in progress (contributors welcome). The MCP + projection layer is device-agnostic; an adapter just maps its source into the same shared schemas, so everything below applies to whatever you wear.

Note: this works through Whoop's private iOS API rather than the public OAuth API. That isn't what Whoop's terms allow — see the FAQ if you want the full picture before installing.

Table of contents

1. Get Started
2. Why this exists
3. What it does
4. Architecture
5. The 55 tools
6. Authentication
7. Write-safety harness
8. Bundled catalogs
9. Configuration
10. Remote hosting
    • Deploying to Hugging Face Spaces (free hosting)
11. The totem CLI
12. Privacy + security
13. Troubleshooting
14. Comparison to alternatives
15. FAQ
16. Disclaimers
17. Acknowledgments

Other root-level docs: TOOLS.md (full per-tool reference) · WHOOP.md (full API reference) · CHANGELOG.md · CONTRIBUTING.md · SECURITY.md · LICENSE

Get Started

Prerequisites: Node 24+ and a Whoop account (any membership tier).

git clone https://github.com/thebriangao/totem.git
cd totem && npm install && npx tsc && npm link

That puts totem on your PATH (the one-time npx tsc builds it; after that every workflow is a totem command — there are no npm run scripts). Now run one guided command — each handles Whoop login (SMS included), setup, and connecting to Claude from end to end. Pick how you want to run it:

★ Recommended — totem cloud

Deploys the server to a host and connects it to Claude on web, desktop, and mobile, synced across every device on your account. One command walks through:

1. Whoop login — email + password, plus the SMS code if your account has MFA. Tokens are saved locally and pushed to the host.
2. Pick a host — Fly (automated + tested), Railway, Google Cloud Run, or your own server (guided Docker steps).
3. Secrets — generates MCP_AUTH_TOKEN and asks you to choose a connector password.
4. Deploy + verify — sets env, deploys, confirms /health and the OAuth endpoints are live.
5. Connect — prints ready-to-paste setup for claude.ai, ChatGPT, Claude Code (remote), and Cursor / Windsurf / any HTTP MCP client (URL + password for the OAuth connectors, a bearer-token block for the rest), and opens claude.ai's connector page.

totem cloud

Local only — totem local

Runs the server on this machine over stdio — no hosting, this device only. Walks through Whoop login → build → wiring it into the client you pick: Claude Desktop, Claude Code, Cursor, VS Code (Copilot), Gemini CLI, Codex CLI, or Windsurf — it writes the right config to the right path automatically (or prints a universal block for any other MCP client). Restart the client and you're set.

totem local

Keeping it alive

Whoop's tokens expire roughly every 30 days. When they do, re-run the same command — it re-logs you in and (if you deployed) pushes the new tokens to your server:

totem auth

Silent if your account has no SMS MFA; prompts for the code if it does. Pushes fresh tokens to your deployment automatically.

Check it works

Ask Claude: "how am I doing today on whoop?" — you should get structured recovery / sleep / strain back.

Prefer to wire everything up by hand? The guided commands just automate the steps documented in The totem CLI, Remote hosting, and Configuration. Stuck? See Troubleshooting.

Why this exists

Whoop ships two APIs:

• The public developer API at developer.whoop.com is OAuth2, read-only, and exposes 13 endpoints under 6 scopes. You get recovery score, sleep stage totals, workout strain, body measurements (3 fields), and HRV/RHR per cycle. No journal, no Strength Trainer, no Whoop Coach, no hypnogram, no stress monitor, no trends, no writes, nothing else. Numeric sport_id was removed 2025-09-01.
• The private iOS API is what the actual Whoop app uses — api.prod.whoop.com behind AWS Cognito. 311 distinct operations across 47 microservices, including everything missing from above.

This MCP wraps the iOS surface.

What the iOS API has that the public OAuth doesn't

If recovery + sleep totals + workout list is enough for you, use the public OAuth API. If anything in the table is interesting, you need this. The iOS API was discovered via mitmproxy — full methodology in WHOOP.md.

What it does

The MCP runs as a local Node process. It speaks Model Context Protocol over stdio (or HTTP for remote deployments), registers 55 tools at startup, and waits for tool calls from a connected MCP client.

When a tool is called:

1. Authenticates via the cached Cognito access token (auto-refreshes if expired)
2. Issues HTTP requests to api.prod.whoop.com
3. Walks the response to extract a flat domain object (the projection step)
4. Validates the projected object against a zod schema (catches Whoop API drift)
5. Returns the structured JSON to the MCP client

Writes follow the same path plus a preview gate: every write tool defaults confirm: false, returning a preview of what would be sent. Claude must explicitly re-call with confirm: true to fire.

See The 55 tools for the full per-tool reference.

Architecture

Claude Desktop / Code  ──stdio──▶  src/server.ts  ──▶  49 tool handlers
                                                          │
                                       ┌──────────────────┼──────────────────┐
                                       ▼                  ▼                  ▼
                                  schemas (zod)    projections (raw→flat)  whoop/client
                                                                              │
                                                                              ▼ HTTPS
                                                                       api.prod.whoop.com

Three layers per tool

Every tool is a schema + projection + handler:

• src/schemas/<tool>.ts — zod schema. The contract Claude sees. Used at runtime to validate the projection's output before returning.
• src/projections/<tool>.ts — pure function turning Whoop's raw BFF response into a flat object. All the "Whoop puts this data over there, not where you'd expect" knowledge lives here. Tested against captured fixtures.
• src/tools/v2/<tool>.ts — ~25-100 lines. Registers the tool, parses input args, calls the client, runs the projection, validates with zod, returns.

Almost no logic in the tool file. That's all in the projection — which makes the codebase highly testable (projections are pure transformations, tested against tests/fixtures/*.json without hitting the network).

Shape drift handling

When Whoop changes a response shape, the projection emits unexpected data, zod's .parse() fails, and the MCP throws WhoopProjectionError instead of silently returning malformed data to Claude. Fix: use whoop_raw + whoop_endpoints to capture the new shape, update the projection, update the fixture, ship.

Recent example: in May 2026, Whoop migrated recovery + strain deep-dives from GRAPHING_CARD tiles (keyed by content.title like "RECOVERY") to SCORE_GAUGE + CONTRIBUTORS_TILE items with stable content.id keys (RECOVERY_SCORE_GAUGE, CONTRIBUTORS_TILE_HRV). Other deep-dives still use the old card-based shape. The escape-hatch tools made the migration trivial to debug.

The 55 tools

Compact summary. Full per-tool reference (input shape · source endpoints · output shape · notes) → TOOLS.md. Tools marked ⚠️ are writes (default confirm: false, preview-first). Tools marked ⚡ are compact/aggregate reads that fold several source calls into one lower-token response; whoop_weekly_plan is ⚡ experimental (its endpoint shape hasn't been verified against a live account yet).

Total: 55 (38 reads + 15 writes + 2 escape hatches). For each tool's input args, source endpoint(s), and output shape, see TOOLS.md.

Authentication

Whoop's iOS app uses AWS Cognito routed through a Whoop-owned proxy (/auth-service/v3/whoop/). The proxy fills in ClientId + SECRET_HASH server-side — no IPA extraction needed.

Bootstrap once (email + password + SMS MFA code if your account has it on) → tokens written to .env. After that, it's hands-off: access tokens auto-refresh every 24h via the refresh token; refresh token lives ~30 days. Single-flight refresh gate prevents thundering-herd refreshes when concurrent tool calls all see a stale token at the same time.

Auth requests impersonate the iOS AWS Swift SDK (the proxy 403s a Node-style User-Agent), and data requests carry the WHOOP iOS app's own identity headers (user-agent: iOS, the x-whoop-* device set, a per-install identifier) so the traffic blends with the legitimate app rather than standing out — see WHOOP.md → Headers for data requests for the full set and the reasoning.

Error classes (src/whoop/errors.ts):

When refresh-token lifetime expires (~30 days), re-run totem auth. It auto-detects local vs deployed and, for a deployment, pushes the new tokens to it. Brand-new SMS code (if your account has MFA), fresh 30-day window.

Write-safety harness

Every write tool defaults confirm: false. The first call returns a preview of what would execute. Claude must explicitly re-call with confirm: true to fire the actual request. Without the gate, a hallucinated "log my workout" could create garbage activities on your account.

The preview shape (lives in src/whoop/write_safety.ts):

{
  "preview": true,
  "will_execute": {
    "method": "POST",
    "path": "/weightlifting-service/v2/weightlifting-workout/activity",
    "body_summary": {
      "exercise_count": 3, "set_count": 12,
      "exercise_list": [{"name": "BENCHPRESS_BARBELL", "set_count": 5}, ...]
    }
  },
  "set_confirm_true_to_run": true
}

Claude reads this back to you, you confirm, Claude re-calls with confirm: true, the actual POST fires. Every write tool's output schema is a withPreview(ReceiptSchema) discriminated union — preview or receipt, never both.

Bundled catalogs

Four datasets compiled into the MCP at build time (not fetched at runtime):

Local lookups, no prompt bloat: the catalog tools are zero-network local searches, so the full 372-exercise / 308-behavior / 203-sport lists stay out of the system prompt (~14k tokens saved). The ID-taking write tools (whoop_activity_create, whoop_lift_log, whoop_journal_log, …) validate their IDs — or an exact catalog name — against these bundled datasets before any request fires, rejecting unknown values with an actionable error. (A session-scoped "must call the catalog first" gate existed through 1.4.4; it was removed in favor of this consistent local validation.)

Configuration

Environment variables

Claude Desktop config

{
  "mcpServers": {
    "whoop": {
      "command": "/opt/homebrew/bin/node",
      "args": ["/absolute/path/to/totem/dist/server.js"]
    }
  }
}

The MCP loads .env from the repo root (relative to server.js). Use absolute paths — Claude Desktop doesn't inherit shell PATH.

Remote hosting

The MCP also speaks HTTP — deploy once, use from multiple devices. Same 55 tools, same auto-refresh, behind a bearer-token gate at a URL.

# 1. Local bootstrap (Cognito needs an interactive MFA prompt)
totem auth

# 2. Build + deploy via the shipped Dockerfile (Fly/Railway/Render/VPS — all work)
docker build -t totem .

# 3. Run with env: WHOOP_EMAIL, WHOOP_IOS_BEARER_TOKEN, WHOOP_COGNITO_REFRESH_TOKEN,
#    MCP_TRANSPORT=http, MCP_AUTH_TOKEN=$(openssl rand -hex 32)

Deploying to Hugging Face Spaces (free hosting)

Hugging Face Spaces provides free Docker hosting — a zero-cost alternative to Fly or Railway. Connecting to claude.ai web and Claude mobile as a custom connector requires a Claude Pro plan, but the server itself runs for free on HuggingFace.

Note: HuggingFace Spaces routes all external traffic through port 7860 only, but this repo's Dockerfile defaults to 3000. The server reads PORT from the environment, so just set PORT=7860 as a Space variable (step 2 below) — no Dockerfile edit needed, and it'll listen on the right port.

Step-by-step

1. Create a new Space

• Go to huggingface.co/new-space
• Select Docker as the Space SDK
• Set the visibility to Public — HuggingFace Spaces must be public for Claude's connector to reach the /mcp endpoint

2. Add your secrets via the Space Settings UI

In your Space → Settings → Variables and Secrets, add the following. The token values go in as Secrets (hidden, never shown in logs); PORT and PUBLIC_URL aren't sensitive, so they can be plain Variables:

Tip: The WHOOP_* token values come from your local .env after running totem auth. MCP_AUTH_TOKEN you generate yourself (openssl rand -hex 32).

3. Push your code to the Space

Clone this repo and add your HuggingFace Space as a remote, then push:

git clone https://github.com/thebriangao/totem
cd totem

# Add your HuggingFace Space as a remote
git remote add hf https://huggingface.co/spaces/<your-hf-username>/<your-space-name>

# Push the main branch
git push hf main

HuggingFace will automatically build and run the Docker container. Watch the Logs tab in your Space to confirm it starts up and /health returns 200.

4. Connect to Claude

Once the Space is running, go to claude.ai → Settings → Connectors → Add custom connector:

• MCP Server URL: https://<your-hf-username>-<your-space-name>.hf.space/mcp
• Claude will open an OAuth flow — enter the AUTH_PASSWORD you set in your Space secrets

MCP_AUTH_TOKEN is used internally as the bearer token + JWT signing secret. You do not type it into Claude's connector UI — that field uses the OAuth AUTH_PASSWORD flow instead.

5. Re-authenticate when tokens expire (~30 days)

Whoop's Cognito tokens expire roughly every 30 days. Re-run totem auth on your local machine, then update WHOOP_IOS_BEARER_TOKEN and WHOOP_COGNITO_REFRESH_TOKEN in your Space's Secrets UI. The Space will restart automatically and pick up the new tokens.

Claude Desktop doesn't natively speak remote MCP — bridge through stdio with mcp-remote:

{ "mcpServers": { "whoop": {
  "command": "npx",
  "args": ["-y", "mcp-remote", "https://your-app.fly.dev/mcp",
           "--header", "Authorization:Bearer your-mcp-auth-token"]
}}}

Claude Code speaks remote MCP natively: claude mcp add whoop --url ... --header ....

claude.ai web + Claude mobile app can't use a bearer token — their custom-connector UI only supports OAuth. The server includes an OAuth 2.1 + PKCE authorization server for exactly this. Enable it by setting two extra env vars:

AUTH_PASSWORD=<a password you'll type once when adding the connector>
PUBLIC_URL=https://your-app.fly.dev   # your server's public origin (the OAuth issuer)

Then in Claude: Settings → Connectors → Add custom connector, paste https://your-app.fly.dev/mcp, and Claude walks the OAuth flow. It pops a small password page (served by your server) — enter AUTH_PASSWORD, approve, done. The connector then syncs across every device logged into your Claude account (web, desktop, mobile). The password gate means a stranger who finds your URL still can't connect without it. MCP_AUTH_TOKEN doubles as the JWT signing secret; leave AUTH_PASSWORD unset to disable this path.

All of the above is what totem cloud automates for you — the manual steps here are for reference or hand-rolling.

When Cognito expires (~30 days): totem auth from your Mac. Silent if your account has no SMS MFA; prompts for the code if it does. Auto-detects your deployment and pushes the new tokens to it (Fly/Railway/Cloud Run/custom), ~10s restart. Requires being at a machine with the repo + the platform CLI.

Security: bearer-token and OAuth paths both gate /mcp. Generate the token random (openssl rand -hex 32), HTTPS only, never commit, rotate if leaked. OAuth access/refresh tokens are stateless signed JWTs (survive restarts); auth codes are one-time + 60s-lived; PKCE S256 is enforced. /health is the only path without auth.

The totem CLI

totem is the single entry point for everything — there are no npm run scripts. It works from any directory (it resolves its own install path), so totem deploy from ~/Desktop does the same thing as cd totem && fly deploy.

# One-time, after cloning:
npm install && npx tsc && npm link   # the only npm steps; `npx tsc` does the first build

totem       # banner + full command list

Commands by group:

Most people only ever need the two Get started commands plus auth (to re-auth when tokens expire). The rest are for power users — totem ping ("is my deploy alive"), totem logs, totem start (drop-in for node dist/server.js), etc.

Staying current — totem update. An interactive updater. Run it and you get a version picker — every release, newest first, with dates and notes pulled from GitHub Releases. The newest is the default; pick anything older and it warns you what you'd be behind on (you can always run it again to go forward). After you choose, it applies the change with a guided, animated build — git installs fast-forward main (or check out the pinned tag), reinstall deps, and rebuild; npm installs reinstall the chosen version — then redeploys to your existing host and health-checks /health. A clean working tree is required (it won't clobber local edits), and a failed deploy leaves your live server untouched.

After a successful update to the latest it offers auto-update. Turn it on and a background job (launchd on macOS, cron on Linux) checks for new releases every 6 hours and applies them the same way — pull, build, redeploy, health-check — logging to .totem-autoupdate.log. There's no push channel to your machine, so it's a scheduled poll, not instant. Toggle it directly anytime:

totem update              # interactive: pick a version, then optionally enable auto-update
totem update --auto on    # turn auto-update on (and pull the newest now)
totem update --auto off   # turn it off
totem update --check       # non-interactive: print installed-vs-latest + auto state

Versions are sourced from the places that already host them — git tags + GitHub Releases for git installs, the npm registry for global installs — so there's nothing extra to maintain. Auto-update is opt-in and only ever moves you to released versions (every release is CI-gated), never bleeding-edge main.

Privacy + security

• Credentials live in .env on your machine. Email, password, access token, refresh token — never leave your filesystem. Claude can't read them (it doesn't have filesystem access unless you wire in a filesystem MCP).
• The only outbound traffic is HTTPS to api.prod.whoop.com. No telemetry, no analytics, no third-party servers. The MCP is open source — every line that touches your data is auditable.
• Write safety: every write tool defaults to confirm: false. The preview shape includes what would be sent. You see it in chat before any mutation. To go further, remove specific writes from src/tools/register.ts or use Claude Desktop's "always require approval" setting.
• Hardened in 1.2.3: token files are written 0600, your account password is wiped from .env after login, OAuth tokens are audience-bound and signed with a key derived from MCP_AUTH_TOKEN, the HTTP server sends anti-clickjacking + CSP headers, the connector-password gate has a global brute-force ceiling, and deploy secrets are pushed over stdin (not argv). Full detail in SECURITY.md and CHANGELOG.md.

Troubleshooting

"AUTH FAIL: Cognito InitiateAuth failed (400)"

Wrong email or password. Double-check .env.

"AUTH FAIL: Cognito MFA challenge missing Session token"

The InitiateAuth response was malformed (unusual). Re-run totem auth — Cognito occasionally drops sessions.

"MFA verification did not return tokens"

You entered the wrong SMS code (or it timed out). Codes expire after ~3 minutes.

"WhoopAuthExpiredError" after every call

Your refresh token has expired (>30 days since last login). Re-run totem auth — it logs you back in, saves the tokens locally, and (if you deployed) auto-detects your host and pushes the new tokens to it in one step. If your account has MFA you'll get a fresh SMS code on your phone to type in the terminal.

"WhoopServerError: 502" / "503" / "504"

Whoop's servers are having issues. Retry in 30 seconds.

Claude says it doesn't see any whoop tools

Check claude_desktop_config.json paths are absolute. Restart Claude Desktop fully (quit, then reopen). Check the MCP server runs without errors: totem dev — it should start silently and wait on stdin.

"WhoopApiError: 422 on /profile-service/v1/profile"

Your whoop_profile_update body is too partial. Send most fields (gender from {MALE,FEMALE,NON_BINARY} only, birthday as YYYY-MM-DD or ISO datetime, country ISO-2). If country=US, also send state — Whoop returns 400 "AdminDivision (state) must be set for US" otherwise.

"Whoop API error 409 on /weightlifting-service/v2/weightlifting-workout/activity"

Time window conflicts with an existing workout. Use a different range.

"WhoopProjectionError for whoop_X"

Whoop changed a response shape. Capture the new response (e.g. via whoop_raw), inspect, update the projection.

Tests fail after git pull

Pull may have updated captured fixtures. Run totem test again to see what changed. If projections need updating, that's the work.

totem build (or npx tsc) fails with "Top-level await is currently not supported with the 'cjs' output format"

You're using an old Node. Upgrade to Node 24+.

"Error: ENOENT: no such file or directory, open '.env'"

Create .env at the repo root (or wherever dist/server.js is being run from — the MCP loads .env relative to the entry).

"Cannot find module '@modelcontextprotocol/sdk/server/mcp.js'"

Run npm install.

"AbortError: This operation was aborted"

A request to Whoop's API took longer than 30s. Either Whoop is slow or your network is slow. Retry.

Comparison to alternatives

This MCP is the only option for programmatic write access to your Whoop data right now.

FAQ

Q: Is this supported by Whoop? A: No. This MCP works through Whoop's private iOS API, which isn't a public surface they intend for third-party tools. Whoop's terms reserve the right to take action against accounts they catch using unsupported integrations — realistically that means suspending API access or terminating the membership. The author has used the MCP heavily for weeks without issue, and traffic patterns look similar to normal app usage, but there's no guarantee. If losing your Whoop account would be a problem for you, don't use this.

Q: Why not use Whoop's public OAuth API instead? A: It's 13 endpoints, all read-only, no journal, no strength, no stress, no coach, no smart alarm, no trends beyond a single recovery score per day. Whoop also pulled numeric sport_id past 2025-09-01 (now sport_name strings only). If you only need recovery score + sleep stage totals + workout list, the OAuth API is the right answer.

Q: Will this work with the Whoop 4.0 vs 5.0 strap? A: Yes — the API doesn't care which strap you have. It cares about your account.

Q: What about Whoop 6.0? A: When it launches and the iOS app updates, the api version may bump from 7 to 8. The MCP's constants.ts may need an update. Worst case, projections break and you fix them.

Q: Can I run this on a server / cloud / always-on? A: Sure. The MCP doesn't care where it runs. Just make sure your .env survives restarts.

Q: Can I share this MCP with my friends? A: Each user needs their own .env with their own Whoop credentials. Don't share tokens.

Q: Is there an HTTP transport instead of stdio? A: Yes — set MCP_TRANSPORT=http to expose the MCP over Streamable HTTP behind a bearer-token + OAuth 2.1 gate (since v1.1.0). See Remote hosting. stdio stays the default for local Claude Desktop / Claude Code.

Q: Does this support Claude's Computer Use API? A: It's MCP-compatible — anything that speaks MCP can talk to it.

Q: Why TypeScript instead of Python? A: The MCP SDK is most mature in TypeScript. Also Whoop's API responses are heavily nested — zod is genuinely the best validation library for that shape work.

Q: Why Node 24 specifically? A: Uses import.meta.dirname (added in 20.11), modern fetch, native ESM, AbortController. Node 18 might work; 16 won't.

Q: How long did this take? A: ~3 weeks of evening/weekend work for v1, plus another week to rewrite as v2 with proper projections and the write-safety harness.

Q: Will you maintain this? A: Best-effort. PRs welcome.

Disclaimers

• Not affiliated with Whoop. "WHOOP" is a trademark of WHOOP, Inc. Community-built tool that interacts with surfaces Whoop has not published. See the FAQ for the practical implications.
• No warranty, use at your own discretion. The API surface is reverse-engineered — Whoop can change response shapes at any time. The zod schemas surface drift as WhoopProjectionError instead of silent corruption.
• Respect rate limits. Single-digit RPS in normal use. Don't be the person who triggers a backend alert that gets every user of this MCP banned.
• Don't share tokens. Your .env is yours. Don't commit it, don't paste it anywhere.

Acknowledgments

• WHOOP for building a fitness platform worth reverse-engineering
• Anthropic for MCP and Claude
• mitmproxy for being the tool that made discovery possible
• The TypeScript + zod community for making strict validation pleasant
• The various API consumers + bloggers who documented bits of Whoop's private API over the years

This is open source under the terms in LICENSE. Contributions welcome.