WHOOP MCP Server

A local Model Context Protocol (MCP) server that gives an LLM read-only access to your WHOOP fitness data. Authentication is direct OAuth against your own WHOOP developer app — there is no third-party proxy in the path. All records are mirrored into a local SQLite cache at ~/.whoop-mcp-server/whoop.db, and no data ever leaves your machine except for the authenticated calls the server itself makes to the WHOOP v2 API.

Current version: 0.8.5 — see CHANGELOG.md.

Pre-1.0 status. The API surface (tool names, response shapes, error codes) is stabilizing but not frozen. Breaking changes may land in 0.x minor bumps during dogfooding. Patch bumps (0.8.x) are bugfix-only. 1.0.0 will be cut when the surface has been stable for 2+ weeks of real use. Pin the minor version in CI if you're building on top of this.

Just want to try it? Copy the prompt in docs/AGENT_INSTALL_PROMPT.md, paste it into Claude Code / Claude Desktop / Cursor / Windsurf, and your agent will do the install end-to-end.

Building on this repo? Start with AGENTS.md and docs/ARCHITECTURE.md — they're the load-bearing conventions and one-page system map.

Table of contents

• What this is

• Install (one line via uvx)

• Install with your agent

• Install from source

• Analysis skill: whoop-insights

• Quick start

• Tool catalog

• MCP resources

• Data model

• Sync model

• Event feed

• Exports

• What's not supported

• Updating

• Operations

• Security and privacy

• Development

• For agents building on this repo

• Versioning

What this is

The WHOOP MCP Server is a small Python process that speaks MCP over stdio. It exposes WHOOP v2 data (profile, body measurement, cycles, recoveries, sleeps, workouts) to any MCP-capable client — primarily Claude Desktop and Claude Code. The server is read-only. It authenticates with WHOOP using an OAuth app you register yourself, so your tokens never travel through a third-party server. The WHOOP records you fetch are written to a local SQLite cache (mode 0o600) so subsequent reads are free and offline, and the cache file never leaves your machine.

Install (one line via uvx)

# 1. Create a WHOOP developer app (one-time, free, self-serve):
#    https://developer-dashboard.whoop.com/apps/create
#    Redirect URI:  http://localhost:8000/callback
#    Scopes:        read:profile read:body_measurement read:recovery
#                   read:cycles read:sleep read:workout offline

# 2. One-shot OAuth — opens browser, catches callback, saves encrypted tokens.
WHOOP_CLIENT_ID="<your client id>" \
WHOOP_CLIENT_SECRET="<your client secret>" \
  uvx --from whoop-mcp whoop-mcp-oauth

# 3. Register with Claude Code. The --env flags are required so the
#    server can refresh tokens when the 1-hour access token expires.
claude mcp add whoop --scope user \
  --env WHOOP_CLIENT_ID="$WHOOP_CLIENT_ID" \
  --env WHOOP_CLIENT_SECRET="$WHOOP_CLIENT_SECRET" \
  -- uvx --from whoop-mcp whoop-mcp

No clone, no venv, no absolute paths. Don't have uv? Install it once: curl -LsSf https://astral.sh/uv/install.sh | sh (or brew install uv).

Install with your agent

Paste the prompt in docs/AGENT_INSTALL_PROMPT.md into any MCP-aware agent (Claude Code, Claude Desktop, Cursor, Windsurf, Zed, Aider). The agent will run the install for you.

Install from source (for development)

git clone https://github.com/AshwanthramKL/whoop-mcp.git
cd whoop-mcp
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt -r requirements-dev.txt
.venv/bin/pytest -q   # 183 tests, ~6s, all respx-mocked

Then register with -- /abs/path/to/whoop-mcp/.venv/bin/python /abs/path/to/whoop-mcp/src/whoop_mcp_server.py.

For Claude Desktop, add the equivalent entry to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS), pointing command at the venv Python and args at src/whoop_mcp_server.py. Run ./scripts/fresh_install_check.sh if you want a clean-env smoke of the install path end-to-end (minus the browser OAuth flow).

Quick start

Once the server is registered, try these prompts:

1. Sync the cache. "Sync my WHOOP data from the last 30 days." Claude calls sync_whoop() and reports per-resource counts.

2. Inspect a day. "Give me yesterday's WHOOP daily summary." Claude calls get_whoop_daily_summary(date="YYYY-MM-DD") and shows a joined cycle + recovery + primary sleep + workouts record.

3. Export. "Export all my cached workouts to ~/whoop-workouts.csv." Claude calls export_whoop(kind="workouts", format="csv", path="...").

Analysis skill: whoop-insights

Bundled with the repo: skills/whoop-insights.

Install (one command):

whoop-mcp-install-skills            # writes to ~/.claude/skills/

That fetches the latest skill from this repo's main branch and drops it where Claude Code (and Cursor / Windsurf / Zed) auto-discovers user-scope skills. Restart your MCP client session and ask:

"How am I doing? Run the whoop-insights skill." "Generate my weekly WHOOP report." "Am I overtraining? Check the last 30 days."

The skill pulls 30 days from the cache, computes personal baselines (HRV, recovery, sleep, strain), flags anomalies with evidence, runs two correlations (sleep→next-day-recovery, strain→next-day-recovery), and optionally generates a self-contained HTML dashboard with Chart.js visualizations. Every claim cites a specific date and a specific number — the skill is instructed never to fabricate.

For repo-native developers iterating on the skill itself: whoop-mcp-install-skills --source ./skills installs from your local working tree instead of GitHub.

Tool catalog

17 tools. All list tools accept start / end as ISO-8601 and auto-paginate. Every read tool takes fresh: bool = False — pass True to bypass the cache and hit the WHOOP API, write-through to cache, and return the live response. Errors are returned as a structured {"error": {...}} envelope; tools never raise.

Error codes: AUTH_FAILED, RATE_LIMITED, NOT_FOUND, UPSTREAM_ERROR, VALIDATION_ERROR, CACHE_ERROR, CACHE_EMPTY, FILE_EXISTS, EXPORT_ERROR, SYNC_ERROR.

MCP resources

The cache is also exposed as read-only MCP resources, so the client can browse date slices without invoking a tool.

All resources return application/json. Bad inputs return the same {"error": {"code","message"}} envelope used by tools.

Data model

Responses are flattened: the WHOOP score wrapper is lifted, milliseconds become seconds (*_seconds, 1 decimal), kilojoules become calories (calories, rounded int), heart-rate keys are renamed to avg_hr_bpm / max_hr_bpm, and sleep stages are named deep_sleep_seconds, rem_sleep_seconds, light_sleep_seconds, awake_seconds, in_bed_seconds. Raw user_id, v1_id, and per-record created_at are dropped.

score_state convention. WHOOP scores are not always computed. When score_state != "SCORED" (e.g. PENDING_SCORE, UNSCORABLE) every derived score field is null and the top-level score_state is preserved so callers know why.

One flattened record per resource:

// Cycle
{
  "id": 123456,
  "start": "2026-04-20T04:00:00.000Z",
  "end": "2026-04-21T04:00:00.000Z",
  "timezone_offset": "+00:00",
  "score_state": "SCORED",
  "strain": 12.4,
  "avg_hr_bpm": 62,
  "max_hr_bpm": 168,
  "calories": 2810
}

// Recovery
{
  "cycle_id": 123456,
  "sleep_id": "bb68db7b-...",
  "score_state": "SCORED",
  "recovery_score": 74,
  "resting_heart_rate_bpm": 48,
  "hrv_rmssd_ms": 87.3,
  "spo2_pct": 97.4,
  "skin_temp_c": 33.1,
  "user_calibrating": false
}

// Sleep
{
  "id": "bb68db7b-...",
  "cycle_id": 123456,
  "start": "2026-04-20T03:10:00.000Z",
  "end": "2026-04-20T10:42:00.000Z",
  "nap": false,
  "score_state": "SCORED",
  "sleep_performance_pct": 88.0,
  "in_bed_seconds": 27120.0,
  "light_sleep_seconds": 12540.0,
  "rem_sleep_seconds": 5280.0,
  "deep_sleep_seconds": 6840.0,
  "awake_seconds": 1080.0
}

// Workout
{
  "id": "a91f...",
  "start": "2026-04-20T17:00:00.000Z",
  "end": "2026-04-20T17:48:00.000Z",
  "sport_name": "Running",
  "score_state": "SCORED",
  "strain": 9.3,
  "avg_hr_bpm": 142,
  "max_hr_bpm": 176,
  "calories": 511,
  "distance_meter": 8030.0,
  "zone_durations_seconds": {
    "zone_zero": 0.0, "zone_one": 120.0, "zone_two": 900.0,
    "zone_three": 1440.0, "zone_four": 420.0, "zone_five": 0.0
  }
}

Sync model

• Cache-first reads. Every list/get tool reads from SQLite by default. An empty window transparently triggers a targeted sync and re-reads.

• fresh=True bypasses the cache, hits the WHOOP API, write-throughs the result into the cache, and returns the live response.

• Incremental sync uses MAX(updated_at) per resource as the cursor. A fresh DB falls back to the last 90 days. sync_whoop(full=True) pulls from 2010-01-01 for every resource (do this once on a brand-new cache).

• Idempotent. Re-running sync_whoop with no new upstream data is a no-op. Combined with snapshot hash dedupe (below), this means the event feed stays quiet.

• Snapshot hash dedupe. profile and body_measurement are singleton snapshots. Before writing, the server compares SHA-256 of the canonical raw payload against the stored blob. Byte-identical payloads do not advance updated_at, so no spurious events are generated.

Event feed

get_whoop_events(since, until=None, resources=None, limit=500) is a chronological feed across all cached resources — pure cache read, no API calls. The window is exclusive on both ends: updated_at > since AND updated_at < until. The strict since bound means you can feed a returned cursor back in as the next since without re-seeing a row.

Each event wraps a flat record:

{"resource": "sleeps",
 "id": "bb68db7b-...",
 "updated_at": "2026-04-20T14:12:33.123Z",
 "record": { "...flat sleep..." }}

Response:

{"status": "success", "count": 17,
 "since": "...", "until": "...",
 "events": [...],
 "next_cursor": null | "<opaque-base64>"}

next_cursor is an opaque base64 encoding of updated_at|resource|id for the last returned event. Pass it back as since to continue. The full triple is used as a tie-broken lower bound, so events sharing an updated_at are never skipped at a pagination boundary. Plain ISO-8601 strings as since still work (M5 compat).

Also exposed as MCP resources — see the MCP resources table above.

Exports

export_whoop(kind, format, path, start=None, end=None, overwrite=False) writes flat cached records to disk. Pure data layer — never hits the API. Run sync_whoop() first.

• kind: cycles | recoveries | sleeps | workouts | all. "all" writes one file per resource (cycles.*, recoveries.*, …) into the given directory.

• format: csv (RFC 4180, alphabetically sorted header, nested values JSON-encoded), jsonl (one record per line, sorted keys), or parquet (pyarrow, snappy).

• overwrite: default False. If the destination has content, returns FILE_EXISTS. True replaces silently.

An empty window still yields a file (header-only CSV / empty JSONL / empty Parquet) so downstream tooling sees a consistent artifact.

What's not supported

Spelled out so nobody wastes an issue.

Updating

MCP servers don't auto-update. When you installed is when you pinned. How to pick up a new release depends on how you installed:

Run health_check() afterwards — its pypi_update_available component will confirm the version your MCP client is now running and whether a newer one is published.

Breaking-change policy: while the project is 0.x, minor bumps (0.8.x → 0.9.0) may include breaking changes — always called out in the ### Changed section of CHANGELOG.md. Patch bumps (0.8.3 → 0.8.5) are bugfix-only. Once we cut 1.0.0, breaking changes require a major bump. Pin to a known-good 0.x.y if you can't absorb churn.

Operations

• Logs. Stderr (always, structured JSON) plus a rotating file at ~/.whoop-mcp-server/logs/whoop-mcp.log (~1 MB per file, 5 backups). Env overrides: WHOOP_LOG_LEVEL (default INFO), WHOOP_LOG_FILE (path; empty string disables the file handler), WHOOP_LOG_JSON (default true).

• health_check(live=True) returns a dict with five component checks (auth, api_reachable, cache_readable, cache_writable, schema_version) plus an overall verdict healthy | degraded | unhealthy. live=False skips the network probe.

• Token refresh. Refresh tokens are used automatically when the access token is within 5 minutes of expiry. An async refresh lock prevents stampedes when multiple in-flight requests discover the same expired token. If refresh fails beyond recovery, re-run whoop-mcp-oauth (after pip install) or .venv/bin/python src/setup_direct_oauth.py (from source).

• Rate limiting. The client respects Retry-After on 429s and retries with exponential backoff. After the retry budget, the call surfaces as RATE_LIMITED — tools never raise.

Security and privacy

Everything runs locally. OAuth tokens are encrypted at rest in ~/.whoop-mcp-server/tokens.json with a key at ~/.whoop-mcp-server/.encryption_key. The SQLite cache file is created with mode 0o600. No third-party services — the only outbound network calls are directly to api.prod.whoop.com. Logs do not include tokens, refresh tokens, the client secret, or raw WHOOP response bodies. See PRIVACY.md for the complete inventory of what the server reads, writes, sends, and logs; and how to delete everything (rm -rf ~/.whoop-mcp-server).

Development

# Run the full test suite (183 tests).
.venv/bin/pytest -q

# Re-record fixtures (live calls, requires creds in env).
.venv/bin/python tests/record_fixtures.py

# Fresh-install smoke test (clones current HEAD into a tempdir and
# verifies the install path end-to-end, minus the browser OAuth flow).
bash scripts/fresh_install_check.sh

Adding a new tool. Add the implementation under src/whoop_mcp_server.py with an @mcp.tool() decorator, a Pydantic model in src/whoop_models.py if the response has a new shape, a cache table in src/whoop_store.py if the resource is persisted, and tests in tests/. Keep the error envelope (_error_payload / _map_error) — tools never raise.

Release flow. Bump src/__version__.py (single source of truth) → add a new top section to CHANGELOG.md in Keep-a-Changelog format → verify SERVER_VERSION in whoop_mcp_server.py picks up the new value (the test_version_import test prevents drift) → update version in pyproject.toml → tag vX.Y.Z on main.

For agents building on this repo

Start here:

• AGENTS.md — load-bearing conventions. File layout, error-envelope discipline, TDD loop, what-not-to-do, release flow. Read before editing anything.

• docs/ARCHITECTURE.md — one-page system map. Four-layer diagram, read/write/sync/event paths, invariants.

• docs/AGENT_INSTALL_PROMPT.md — the copy-pasteable prompt end users give their agent to install.

Human contributors: CONTRIBUTING.md has the short version. The test suite runs in ~6s (.venv/bin/pytest -q); every HTTP call is respx-mocked so you can iterate offline.

Versioning

Current version: 0.8.5 (see src/__version__.py). Semantic versioning. Full history: CHANGELOG.md.

Credits

Forked from RomanEvstigneev/whoop-mcp-server (v0.1.x). Substantially rewritten starting at v0.2.0 to use direct WHOOP OAuth (no third-party proxy), the WHOOP v2 API, Pydantic-flattened responses, a local SQLite cache with incremental sync, cache-first reads, an event feed, exports (CSV/JSONL/Parquet), a health_check tool, structured JSON logging, and 190+ tests (190 as of 0.8.5). Licensed MIT — see LICENSE for both copyright lines.

mcp-name: io.github.AshwanthramKL/whoop-mcp