Strava + Calendar¶

The second functional path pairs a read-only Strava metrics source with a calendar as the workout-calendar source (Google Calendar is the lead binding; Outlook Calendar is a second, equivalent binding). See Capabilities & paths for the full capability comparison — garmin covers nine of twelve capabilities, this path covers three (activity_streams, prs, free_text_workouts) plus a derived training_load. Below are the concrete tool inventories and auth steps for each binding.

Strava (metrics) — read-only¶

Reuses the repo's existing strava_server.py / strava_client.py (or the community r-huijts/strava-mcp, TS/stdio) behind a curated allowlist, the same pattern as Garmin's GARMIN_ENABLED_TOOLS:

Tool	Returns
`get-activities`	Recent activity list (curated summary — type, duration, distance, avg HR/pace/power)
`get-activity-streams`	Time-series HR / power / pace / cadence for a single activity
`get-athlete-stats`	Rolling totals (recent/YTD/all-time distance, elevation, ride/run counts)
`get-segment-prs`	Personal-record efforts on starred/visited segments
`get-athlete-zones`	HR and power zone boundaries, used to bucket streams without a training-effect score

Read-only by design

Strava's API has no endpoint for planned or structured workouts — there is nothing to write. This is why strava only ever carries the metrics role, not workout_calendar. A calendar source supplies workout_calendar. Auth reuses scripts/setup_auth.py's existing OAuth flow.

Google Calendar (workout-calendar) — lead binding¶

Uses nspady/google-calendar-mcp (TS/stdio). A workout is a calendar event; the session plan (warm-up, intervals, target paces/HR, notes) lives in the event's free-text description.

Tool	Use in Coach AI
`list-events`	Find today's/this week's workout events (= `get_scheduled_workouts` equivalent)
`get-event`	Read a single event's `description` for the planned session
`create-event`	Write a new workout — title + `description` holds the structured plan (= `create_* → schedule_workout` equivalent)
`update-event`	Modify in place — rewrite `description`/time for `adjust-workout` (= Garmin `unschedule_workout` → recreate → reschedule, but one call)
`delete-event`	Remove a workout event (rarely used — `update-event` is preferred for adjustments)

Auth: desktop OAuth via a gcp-oauth.keys.json credentials file (Google Cloud Console → OAuth client, "Desktop app" type); the MCP server runs a local browser-based consent flow once and caches the refresh token.

Google OAuth 'testing' mode

While the OAuth consent screen is in "testing" status, refresh tokens expire after 7 days, silently breaking the calendar connection until you re-auth. For personal long-term use, publish the consent screen (or move it to internal/production status) — see Roadmap for the full risk note.

Outlook Calendar (workout-calendar) — second binding¶

Uses softeria/ms-365-mcp-server (TS/stdio) with the same event-as-workout model — the session plan lives in the event's body field.

Tool	Use in Coach AI
`list-events`	Find today's/this week's workout events
`get-event`	Read a single event's `body` for the planned session
`create-event`	Write a new workout — title + `body` holds the structured plan
`update-event`	Modify in place (PATCH) — rewrite `body`/time for `adjust-workout`

Auth: MSAL device-code flow against an Azure app registration with the Calendars.ReadWrite delegated scope — more setup overhead than Google's (an Azure tenant app registration is required).

Workout = calendar event, modify = update-event¶

flowchart TD
    A["create-event(title, start, end, description/body = workout plan)<br/>Google or Outlook"] -->|"returns event_id"| B["PlanNote.workout_ref = event_id<br/>workout_source = 'google_calendar' / 'outlook_calendar'"]
    B -->|"on adjust-workout"| C["update-event(event_id, new description/body, new time)<br/>one call — no unschedule/recreate cycle"]

This is simpler than the Garmin modify path (unschedule_workout → create_* → schedule_workout) precisely because a calendar event is a single mutable object — but it puts more weight on the agent to keep the description/body text well-structured, since there's no builder schema enforcing it. generate-daily-workout and adjust-workout write the description in a fixed lightweight template, and the agent re-parses its own template on read — no external schema needed.