coros-mcp

A Model Context Protocol (MCP) server that fetches sleep, HRV, and training data from the unofficial Coros API and exposes them to AI assistants like Claude.

No API key required. This server authenticates directly with your Coros Training Hub credentials. Your token is stored securely in your system keyring (or an encrypted local file as fallback), never transmitted anywhere except to Coros.

What You Can Do

Ask your AI assistant questions like:

• "How much deep sleep and REM did I get last week?"

• "What was my HRV trend over the last 4 weeks?"

• "Show me my resting heart rate and training load for last week"

• "How many steps did I average per day this month?"

• "List my rides from last month"

• "Show me the details of my last long ride"

• "Create a 90-minute sweet spot workout for me"

• "What's on my training calendar next week?"

• "Schedule my VO2 workout for Thursday"

• "Create a 20-minute strength circuit with squats, lunges, and planks"

Related MCP server: whoop-mcp

Features

Setup

Option A: Auto-Setup with Claude Code

If you have Claude Code, paste this prompt:

Set up the Coros MCP server from https://github.com/cygnusb/coros-mcp — clone it, create a venv, install it with pip install -e ., add it to my MCP config, then tell me to run 'coros-mcp auth' in my terminal to authenticate.

Claude will handle the installation and guide you through configuration.

Option B: Manual Setup

Step 1: Install

git clone https://github.com/cygnusb/coros-mcp.git
cd coros-mcp
python3 -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -e .

Or with uv:

uv pip install -e .

Step 2: Add to Claude Code

claude mcp add coros -- /path/to/coros-mcp/.venv/bin/coros-mcp serve

To limit the MCP to a specific project only (recommended):

cd /path/to/your/project
claude mcp add --scope project coros -- /path/to/coros-mcp/.venv/bin/coros-mcp serve

Or add to Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "coros": {
      "command": "/path/to/coros-mcp/.venv/bin/coros-mcp",
      "args": ["serve"]
    }
  }
}

Step 3: Authenticate

Option A — .env file (recommended for project-scoped setups):

Create a .env file in your project directory:

COROS_EMAIL=you@example.com
COROS_PASSWORD=yourpassword
COROS_REGION=eu

The server authenticates automatically on the first request and re-authenticates transparently whenever the token expires. No manual auth step needed.

Option B — Manual authentication:

Run the following command in your terminal — outside of any Claude session:

coros-mcp auth

You will be prompted for your email, password, and region (eu, us, or asia). This stores both the Training Hub web token and the mobile API token (used for sleep data). Your credentials are sent directly to Coros and the tokens are stored securely in your system keyring (or an encrypted local file as fallback). You only need to do this once — the tokens persist across restarts.

Note: The mobile login (apieu.coros.com) will log you out of the Coros mobile app on your phone. If you want to avoid this, use coros-mcp auth-web instead — it stores only the web token, and the mobile token will be obtained automatically when you first request sleep data.

Other auth commands:

coros-mcp serve         # Start the MCP server (used by Claude Code / Claude Desktop)
coros-mcp auth-web      # Web API only — skips mobile login (sleep data obtained lazily)
coros-mcp auth-mobile   # Mobile API only (sleep data)
coros-mcp auth-status   # Check if authenticated
coros-mcp auth-clear    # Remove stored tokens

Step 4 (optional): Restrict the toolset for autonomous agents

When the MCP client is an autonomous agent or a smaller open-weight model (e.g. hermes-agent with a DeepSeek/Hermes model), two env vars harden the server:

COROS_MCP_TOOLSET=readonly       # expose only the 12 read tools
COROS_MCP_HIDE_AUTH_TOOLS=1      # hide authenticate_* (credentials never enter the model context)

• COROS_MCP_TOOLSET=readonly hides everything that writes to your Coros account (save/schedule/update/delete tools) plus the low-level raw escape hatches, leaving: get_help, check_coros_auth, get_daily_metrics, get_sleep_data, list_activities, get_activity_detail, list_workout_templates, list_training_plans, list_planned_activities, list_exercises, sync_coros_data, get_cache_status. A smaller tool list also improves tool-calling reliability of non-frontier models. Default: full (all 26 tools).

• COROS_MCP_HIDE_AUTH_TOOLS=1 removes the two authenticate_* tools so your password can never travel through the model context or agent logs. Authentication then happens exclusively via the COROS_EMAIL / COROS_PASSWORD env auto-login (or tokens stored beforehand with coros-mcp auth).

All tools additionally carry MCP readOnlyHint / destructiveHint annotations, so agent harnesses that honor them can require confirmation for write actions.

Tool Reference

authenticate_coros

Log in with your Coros credentials. The auth token is stored securely in your system keyring (or an encrypted file as fallback).

{ "email": "you@example.com", "password": "yourpassword", "region": "eu" }

Returns: authenticated, user_id, region, message

authenticate_coros_mobile

Authenticate with the Coros mobile API only. This is mainly useful if you need to restore sleep-data access without redoing full web authentication.

{ "email": "you@example.com", "password": "yourpassword", "region": "eu" }

Returns: authenticated, user_id, region, message

check_coros_auth

Check whether valid web and mobile tokens are stored and how long the web token remains valid.

{}

Returns: authenticated, user_id, region, expires_in_hours, mobile_authenticated, mobile_token_status

get_daily_metrics

Fetch daily metrics for a configurable number of weeks (default: 4).

{ "weeks": 4 }

Returns: records (list), count, date_range

Each record includes:

get_sleep_data

Fetch nightly sleep stage data for a configurable number of weeks (default: 4).

{ "weeks": 4 }

Returns: records (list), count, date_range

Each record includes:

Note: Sleep data is fetched from the Coros mobile API (apieu.coros.com), which uses a separate token from the Training Hub web API. coros-mcp auth obtains both tokens, but doing so logs you out of the Coros mobile app. Use coros-mcp auth-web to skip mobile login — the mobile token is then obtained automatically on the first sleep data request. The token expires after ~1 hour but refreshes automatically on subsequent requests.

list_activities

List activities for a date range.

{ "start_day": "20260101", "end_day": "20260305", "page": 1, "size": 30 }

Returns: activities (list), total_count, page

Each activity includes: activity_id, name, sport_type, sport_name, start_time, end_time, duration_seconds, distance_meters, avg_hr, max_hr, calories (in cal — divide by 1000 for kcal), training_load, avg_power, normalized_power, elevation_gain

get_activity_detail

Fetch full detail for a single activity. Requires the sport_type from list_activities.

{ "activity_id": "469901014965714948", "sport_type": 200 }

Returns full activity data including laps, HR zones, power zones, and all sport-specific metrics.

Note: Large time-series arrays (graphList, frequencyList, gpsLightDuration) are stripped from the response to keep it manageable.

list_workout_templates

List reusable workout templates saved in the Coros library.

{}

Returns: workouts (list), count

Each entry includes: id, name, sport_type, sport_name, estimated_time_seconds, exercise_count, exercises (list of steps with name, intensity_low, intensity_high, sets, and exactly one duration key: duration_seconds, distance_meters, or duration_open)

save_workout_template

Save a reusable cycling/intervals workout template to the Coros library. The template appears in the Coros app and can be synced to the watch. Steps can be plain steps or repeat groups for intervals.

⚠️ This persists to the library indefinitely. For a one-off workout for a specific date, use schedule_workout instead — it builds the workout inline and leaves no library entry.

Plain steps:

{
  "name": "Sweet Spot 90min",
  "sport_type": 2,
  "steps": [
    {"name": "15:00 Warmup",     "duration_minutes": 15, "intensity_low": 148, "intensity_high": 192},
    {"name": "20:00 Sweet Spot", "duration_minutes": 20, "intensity_low": 260, "intensity_high": 275},
    {"name": "5:00 Rest",        "duration_minutes":  5, "intensity_low": 100, "intensity_high": 150},
    {"name": "20:00 Sweet Spot", "duration_minutes": 20, "intensity_low": 260, "intensity_high": 275},
    {"name": "30:00 Cooldown",   "duration_minutes": 30, "intensity_low": 100, "intensity_high": 192}
  ]
}

With repeat groups (intervals):

{
  "name": "3×10min Sweet Spot",
  "sport_type": 2,
  "steps": [
    {"name": "Warmup",    "duration_minutes": 10, "intensity_low": 150, "intensity_high": 200},
    {"repeat": 3, "steps": [
      {"name": "Sweet Spot", "duration_minutes": 10, "intensity_low": 265, "intensity_high": 285},
      {"name": "Recovery",   "duration_minutes":  3, "intensity_low": 150, "intensity_high": 175}
    ]},
    {"name": "Cooldown",  "duration_minutes": 11, "intensity_low": 150, "intensity_high": 200}
  ]
}

Distance-based steps: use duration_meters instead of duration_minutes for a step that should end at a real distance regardless of pace, e.g. {"name": "1km @ 4:00/km", "duration_meters": 1000, "intensity_low": 235, "intensity_high": 245} (pace in sec/km with intensity_type: 3).

Open (lap-press) steps: use duration_open: true for a step with no clock and no distance cap — the watch shows no countdown and advances only when the athlete presses lap, e.g. {"name": "Warm-up", "duration_open": true, "intensity_low": 120, "intensity_high": 150}. Intensity bounds still show as a guide zone but never end the step. Because an open step's length is unknowable ahead of time, it counts toward neither total_minutes nor distance_meters_total, and the response carries a warning saying so.

Exactly one of the three keys (duration_minutes, duration_meters, duration_open) per step.

sport_type: 2 = Indoor Cycling (default), 200 = Road Bike

Returns: workout_id, name, total_minutes, distance_meters_total, steps_count, message

save_strength_workout_template

Save a reusable strength workout template to the Coros library. Exercises must come from the Coros exercise catalogue.

⚠️ This persists to the library indefinitely. For a one-off workout for a specific date, use schedule_strength_workout instead — it builds the workout inline and leaves no library entry.

{
  "name": "Leg Circuit",
  "sets": 3,
  "exercises": [
    {
      "origin_id": "54",
      "name": "T1061",
      "overview": "sid_strength_squats",
      "target_type": 3,
      "target_value": 12,
      "rest_seconds": 45,
      "sets": 3,
      "weight_kg": 80
    },
    {
      "origin_id": "130",
      "name": "T1176",
      "overview": "sid_strength_plank",
      "target_type": 2,
      "target_value": 60,
      "rest_seconds": 30
    }
  ]
}

target_type: 2 = time in seconds, 3 = reps

sets (per exercise, optional): consecutive sets of that exercise before moving on. Use this instead of duplicating the exercise entry. Defaults to 1.

sets (top-level): full-circuit repetitions — repeats the entire exercise list. Defaults to 1.

Weight fields (per exercise, optional):

• Omitting both fields renders the exercise as Bodyweight in the Coros app.

• Setting weight_kg: 0 renders as "0.00 kg" — distinct from Bodyweight.

• For dumbbell exercises, the weight is per hand by convention (the app shows a single value).

• kg and lbs can be mixed across exercises within the same workout.

Returns: workout_id, name, sets, exercise_count

delete_workout_template

Delete a saved workout template from the Coros library.

{ "workout_id": "476023839273435149" }

The workout_id comes from list_workout_templates.

Returns: deleted, workout_id, message

list_planned_activities

List planned activities from the Coros training calendar.

{ "start_day": "20260309", "end_day": "20260316" }

Returns: schedule (dict with entities and programs sub-lists), count (number of scheduled entities), date_range

schedule_workout

Schedule a one-off cycling/intervals workout for a specific date. Builds the workout inline and posts straight to the calendar — does not create a library entry. This is the common case.

Same parameters as save_workout_template, plus happen_day (YYYYMMDD) and an optional sort_no (order within the day).

{
  "name": "Z2 Recovery",
  "happen_day": "20260312",
  "sport_type": 2,
  "steps": [
    {"name": "Easy spin", "duration_minutes": 60, "intensity_low": 150, "intensity_high": 200}
  ]
}

Returns: scheduled, name, happen_day, total_minutes, steps_count, response

schedule_strength_workout

Schedule a one-off strength workout for a specific date. Builds the workout inline and posts straight to the calendar — does not create a library entry.

Same parameters as save_strength_workout_template, plus happen_day.

{
  "name": "Push Day",
  "happen_day": "20260312",
  "sets": 1,
  "exercises": [ ... ]
}

Returns: scheduled, name, happen_day, sets, exercise_count, response

schedule_workout_template

Schedule an existing library template on a calendar day.

{ "workout_id": "1234567890", "happen_day": "20260312", "sort_no": 1 }

The workout_id comes from list_workout_templates. For one-off workouts that don't need a library entry, use schedule_workout or schedule_strength_workout instead.

Returns: scheduled, workout_id, happen_day, response

remove_scheduled_workout

Remove a scheduled workout from the Coros training calendar.

{
  "plan_id": "987654321",
  "id_in_plan": "1234567890",
  "plan_program_id": "1234567890"
}

plan_id, id_in_plan, and plan_program_id come from list_planned_activities. If plan_program_id is missing, you can usually reuse id_in_plan.

Returns: removed, plan_id, id_in_plan

list_exercises

List the Coros exercise catalogue for a sport type. Default sport_type=4 returns strength exercises.

{ "sport_type": 4 }

Returns: exercises (list), count, sport_type

sync_coros_data

Backfill all data into the local SQLite cache for a date range. After the first full sync, get_daily_metrics, get_sleep_data, and list_activities serve historical data from cache and only fetch the incremental tail from the API.

{ "start_day": "20230101", "end_day": "20260514" }

Both parameters are optional and default to two years ago / today respectively. For large ranges (> 6 months) prefer the CLI:

coros-mcp sync --from 20230101

Returns: daily (records synced), sleep (records synced), activities (records synced), errors (list), cache (coverage summary)

get_cache_status

Show what data is currently stored in the local SQLite cache.

{}

Returns per data type: count, from (earliest date), to (latest date). Also includes db_path.

Requirements

• Python ≥ 3.11

• A Coros account (Training Hub)

Project Structure

coros-mcp/
├── coros_mcp/
│   ├── server.py      # MCP server with tool definitions
│   ├── coros_api.py   # Coros API client (auth, requests, parsers)
│   ├── models.py      # Pydantic data models
│   ├── cli.py         # CLI entry point (serve, auth, sync, cache-status, …)
│   ├── auth/          # Token storage (keyring + encrypted file fallback)
│   └── cache/         # SQLite cache layer (store, sync, utils)
├── tests/             # pytest test suite
├── pyproject.toml     # Project metadata & dependencies
└── docs/
    └── mobile-token.md  # Mobile API token background (legacy reference)

Dependencies

• fastmcp — MCP framework

• httpx — Async HTTP client

• pydantic — Data validation

• cryptography — AES encryption (mobile API auth + token file)

• keyring — Secure token storage

• python-dotenv — .env support

Disclaimer

This project uses the unofficial Coros Training Hub API. The API may change at any time without notice. Use at your own risk.