coros-training-mcp

Running-first MCP server for COROS. Author, edit, and schedule running workouts — pace-based intervals, distance targets, repeat groups, clone-and-swap edits — plus sleep, HRV, training load, and activity exports.

Runs locally as a stdio subprocess of your AI assistant. No API key required, no public endpoint. Credentials live in your OS keyring; traffic is outbound only, directly to COROS.

Landing page & screenshots: https://dholliday3.github.io/coros-training-mcp/

Install

uv tool install coros-training-mcp
coros-mcp setup

(pipx install coros-training-mcp works too.) The wizard asks for your COROS email, password, and region, verifies them against the API, stores them in your system keyring, detects which AI assistants you have installed (Claude Code, Claude Desktop, Codex CLI, Cursor), writes the MCP entry for each one you pick (preserving any existing entries), and runs a smoke test.

Requirements: Python ≥ 3.11 (uv tool install fetches it automatically), a COROS Training Hub account, macOS / Linux / Windows.

Lifecycle:

coros-mcp setup --reconfigure   # change credentials or add more assistants
coros-mcp uninstall             # remove from assistants, optionally clear keyring
coros-mcp auth-status           # check stored tokens
uv tool upgrade coros-training-mcp

Related MCP server: garmin-coach-mcp

What you can ask

Running workouts (the focus):

• "Create a 5×1km threshold workout at 4:05–4:15/km with 2-minute jog recovery"

• "Change my Tuesday VO2 workout to 6 reps instead of 5"

• "Build a 90-minute long run with 4×8min at marathon pace in the middle"

• "Move Thursday's tempo run to Friday"

• "Replace my scheduled Sunday long run with a 16km progression at easy→steady"

Recovery & training data:

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

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

• "Show me resting HR and training load for last week"

Activities, schedule, and other sports:

• "List my rides from last month"

• "Export my Saturday long run to GPX"

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

• "Create a 90-minute sweet spot workout"

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

Tools

Workout taxonomy

Three distinct objects that are easy to confuse:

• Library workout — reusable program in your account. Queried by list_workouts.

• Scheduled entry — calendar occurrence of a workout on a specific day. Queried by list_scheduled_workouts. Has different IDs than its source workout.

• Plan container — higher-level training plan that owns scheduled entries.

Full detail: docs/workout-taxonomy.md. Enum-extraction mechanics: docs/enum-extraction.md.

Export (GPX / FIT / TCX / KML / CSV) applies to completed activities only — structured library workouts use a separate share flow in the COROS app. The MCP writes to COROS server state; device sync is still handled by the app/watch.

Privacy & data handling

• Credentials: system keyring (macOS Keychain / Windows Credential Manager / freedesktop Secret Service). If the keyring is unavailable (headless Linux, some VMs), the wizard falls back to an AES-encrypted file at ~/.coros-mcp/credentials.enc and tells you.

• Assistant config entries: live in each assistant's own config file. Only a coros entry is added or replaced; other MCP entries are never touched.

• Network: outbound TLS only, directly to teamapi.coros.com / teameuapi.coros.com / apieu.coros.com. No telemetry, no analytics, no third-party services.

• Binary: coros-mcp lives in the uv tool / pipx isolated venv at an absolute path that MCP clients pin to.

Tool reference

Auth — authenticate_coros, authenticate_coros_mobile, check_coros_auth

You normally don't call these directly — credentials from the keyring or .env are picked up automatically. They exist for explicit login/reauth ({ "email", "password", "region" }) and status checks (check_coros_auth returns authenticated, expires_in_hours, mobile_authenticated, mobile_token_status). authenticate_coros_mobile is useful for restoring sleep-data access without redoing web auth.

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, count, date_range. Each record includes date, total_duration_minutes, phases.{deep,light,rem,awake,nap}_minutes, avg_hr, min_hr, max_hr, quality_score.

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, but doing so logs you out of the COROS mobile app on your phone. Use coros-mcp auth-web (or let the wizard's default skip-mobile choice stand) — the mobile token is then fetched lazily on the first sleep-data request and refreshed automatically.

list_activities

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

Returns: activities, 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, training_load, avg_power, normalized_power, elevation_gain.

get_activity_detail

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

Full activity data including laps, HR zones, power zones, and sport-specific metrics. Large time-series arrays (graphList, frequencyList, gpsLightDuration) are stripped to keep the response manageable.

export_activity_file

{
  "activity_id": "469901014965714948",
  "sport_type": 100,
  "file_type": "gpx",
  "output_path": "/tmp/morning-run.gpx"
}

file_type: gpx, fit, tcx, kml, or csv. Returns activity_id, sport_type, file_type, file_url, output_path, downloaded.

list_workouts

{}

Returns workouts, count. Each workout includes id, name, sport_type, sport_name, estimated_time_seconds, exercise_count, exercises (steps with name, duration_seconds, power_low_w, power_high_w).

create_run_workout

Run-native step kinds (warmup, training, rest, cooldown), distance / time / open targets, optional pace / HR intensity ranges, nested repeat groups.

{
  "name": "Tuesday Threshold",
  "steps": [
    {"kind": "warmup",   "name": "Warm-up",  "target_type": "distance", "target_distance_meters": 2000},
    {"repeat": 5, "name": "Main Set", "steps": [
      {"kind": "training", "name": "Rep",      "target_type": "distance", "target_distance_meters": 1000,
       "intensity_type": 3, "intensity_value": 245, "intensity_value_extend": 255, "intensity_display_unit": 2},
      {"kind": "rest",     "name": "Recovery", "target_type": "time",     "target_duration_seconds": 120}
    ]},
    {"kind": "cooldown", "name": "Cool-down", "target_type": "distance", "target_distance_meters": 1500}
  ]
}

Pace targets use intensity_type: 3 with intensity_value / intensity_value_extend as seconds-per-km and intensity_display_unit: 2. Friendly pace strings like "4:05-4:15/km" or "5:30/mi" are also accepted on any run step as a "pace" field. For the full field vocabulary (HR zones, percent-of-LT, named intensity presets) call get_run_workout_schema or see run_workout_schema.py.

Open steps. "target_type": "open" is the Training Hub "Open" target (COROS targetType=1, manualEnd): the step has no target and runs until the athlete presses the lap button. Omit target_duration_seconds / target_distance_meters — they're ignored. This is how you author an open warm-up or cool-down so the athlete, not the watch, decides the length:

{
  "name": "Sub-T 3×6min",
  "steps": [
    {"kind": "warmup", "name": "WU open", "target_type": "open"},
    {"repeat": 3, "steps": [
      {"kind": "training", "name": "Sub-T 6min", "target_type": "time", "target_duration_seconds": 360,
       "intensity_label": "Heart Rate", "intensity_value": 168, "intensity_value_extend": 175},
      {"kind": "rest", "name": "Jog", "target_type": "time", "target_duration_seconds": 120}
    ]},
    {"kind": "cooldown", "name": "CD open", "target_type": "open"}
  ]
}

Open steps contribute nothing to the workout's estimated time/distance — only the targeted steps do. "manual" and "lap" are accepted aliases. Works the same on create_bike_workout, and on the update_* patch path (switching a step to open clears its stale target).

Returns: workout_id, sport_type, estimated_time_seconds, estimated_distance_meters, steps_count, message.

update_run_workout

Clone-and-edit an existing running workout. Select each step by step_name, step_id, or step_index and patch it with any run-step field used by create_run_workout. Original is preserved; a new workout ID is returned. If the original was scheduled, use replace_scheduled_workout (or schedule_workout) to swap the calendar entry.

{
  "workout_id": "476023839273435149",
  "name": "Tuesday Threshold (6×1km)",
  "step_updates": [
    {"step_name": "Main Set", "repeat": 6},
    {"step_name": "Rep", "target_distance_meters": 1000, "intensity_value": 240, "intensity_value_extend": 250}
  ]
}

Returns: new_workout_id, original_workout_id, name, steps_count, message.

get_run_workout_schema

Returns the shared schema used by create_run_workout and update_run_workout: allowed step kinds, target types, intensity presets pulled from the live Training Hub builder, and required vs. optional fields. Call this before authoring to avoid guessing.

create_bike_workout

Bike-native step kinds (warmup, training, rest, cooldown), time / distance / open targets, optional power / HR / cadence / FTP-percent intensity, nested repeat groups. Sets sportType=2 so warm-up / training / rest / cool-down segments render correctly on the watch (unlike create_workout, which builds flat training blocks).

{
  "name": "3×15 Sweetspot",
  "steps": [
    {"kind": "warmup",   "name": "Warm-up",  "target_type": "time", "target_duration_seconds": 600, "power": "120-160W"},
    {"repeat": 3, "name": "Main Set", "steps": [
      {"kind": "training", "name": "Sweetspot", "target_type": "time", "target_duration_seconds": 900, "power": "230-260W"},
      {"kind": "rest",     "name": "Recovery",  "target_type": "time", "target_duration_seconds": 180}
    ]},
    {"kind": "cooldown", "name": "Cool-down", "target_type": "time", "target_duration_seconds": 600}
  ]
}

Power targets use intensity_type: 6 with intensity_value / intensity_value_extend in watts. Friendly strings like "220W" or "220-260W" are also accepted on any bike step as a "power" field. Raw intensity fields work for HR, cadence, FTP-percent, and speed targets. For the full field vocabulary call get_bike_workout_schema or see bike_workout_schema.py.

Returns: workout_id, sport_type, sport_name, estimated_time_seconds, estimated_distance_meters, steps_count, message.

update_bike_workout

Clone-and-edit an existing cycling workout. Select each step by step_name, step_id, or step_index and patch it with any bike-step field used by create_bike_workout. Original is preserved; a new workout ID is returned. If the original was scheduled, use replace_scheduled_workout (or schedule_workout) to swap the calendar entry.

{
  "workout_id": "476023839273435149",
  "name": "4×15 Sweetspot",
  "step_updates": [
    {"step_name": "Sweetspot", "target_duration_seconds": 900, "power": "240-270W"}
  ]
}

get_bike_workout_schema

Returns the shared schema used by create_bike_workout and update_bike_workout. Intensity label presets (Power, %FTP, Heart Rate, Cadence, Speed, etc.) are pulled from the checked-in live Training Hub builder catalog. If the catalog hasn't yet been recaptured to include bike correlations, the preset list will be empty and you should pass the raw intensity_* fields or use the power ergonomic directly.

create_workout

Generic time-and-power builder (legacy cycling path). For cycling, prefer create_bike_workout — it understands segment kinds (warm-up / training / rest / cool-down) and renders them correctly on the watch. For running, use create_run_workout.

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

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

update_workout

Lower-level clone-and-patch primitive used by update_run_workout and for strength edits. Patch steps/exercises by step_name, step_id, step_index, or origin_id (strength exercise swap). Supports rest_seconds, target_type (time|reps|distance), and any field the create side accepts.

delete_workout

{ "workout_id": "476023839273435149" }

list_planned_activities / list_scheduled_workouts

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

Returns scheduled entries for the window, including library-sourced and plan-embedded programs. Use list_scheduled_workouts for the canonical MCP-facing shape.

schedule_workout

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

move_scheduled_workout

Move a scheduled entry to a new day without losing the underlying workout. Handles both library-sourced entries and plan-embedded programs (which have no library counterpart).

{ "plan_id": "987654321", "id_in_plan": "1234567890", "new_happen_day": "20260314" }

replace_scheduled_workout

Swap a scheduled entry for a different workout (typically a freshly updated clone) in-place. Preserves the calendar slot and sort order.

{ "plan_id": "987654321", "id_in_plan": "1234567890", "replacement_workout_id": "476023839273435149" }

remove_scheduled_workout

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

If plan_program_id is missing from list_planned_activities, reuse id_in_plan.

create_strength_workout

Structured strength workout with repeated sets. Exercises come from the COROS catalog (list_exercises).

{
  "name": "Leg Circuit",
  "sets": 3,
  "exercises": [
    {"origin_id": "54",  "name": "T1061", "overview": "sid_strength_squats", "target_type": 3, "target_value": 12, "rest_seconds": 45},
    {"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.

list_exercises

{ "sport_type": 4 }

sport_type=4 is strength. Returns exercises, count, sport_type.

Manual setup (advanced)

If you're not using one of the auto-detected assistants, install the server and point any MCP client at it:

uv tool install coros-training-mcp
coros-mcp auth          # interactive login, stores tokens in keyring
which coros-mcp         # absolute path for the config below

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

Developer setup

git clone https://github.com/dholliday3/coros-training-mcp.git
cd coros-training-mcp
python3 -m venv .venv && source .venv/bin/activate
pip install -e .[dev]
pytest

CLI reference

coros-mcp setup                 # first-time interactive setup
coros-mcp setup --reconfigure   # re-run wizard
coros-mcp uninstall             # remove from assistants
coros-mcp serve                 # start the MCP server (what MCP clients run)
coros-mcp auth                  # re-authenticate (web + mobile)
coros-mcp auth-web              # web token only (sleep data lazy-loads)
coros-mcp auth-mobile           # mobile token only
coros-mcp auth-status           # check stored tokens
coros-mcp auth-clear            # remove stored tokens

Built on top of cygnusb/coros-mcp, kept as an upstream reference.

Disclaimer

Uses the unofficial COROS Training Hub API. The API may change at any time without notice. Not affiliated with or endorsed by COROS. Use at your own risk.