Run Coach

An MCP server that pulls your Strava data and gives you a run recommendation from Claude, rendered as a dashboard.

What it does

• Fetches recent activities and all-time stats from the Strava API
• Shows them in a Prefab UI dashboard (rides, runs, distance, suffering index)
• Has a button that sends your last N days of runs to Claude and gets back a plain-English recommendation for today's run

Stack

• FastMCP — the MCP server itself, with FastMCPApp for the UI-facing tools
• Prefab — the dashboard UI, written in Python instead of JSX
• Anthropic API — generates the run recommendation
• httpx — talks to Strava
• uv — dependency management and running the thing

Setup

1. Install dependencies

uv add fastmcp anthropic httpx

2. Strava API access

You need a Strava API app (create one at strava.com/settings/api) and a one-time OAuth flow to get a refresh token. Scope needs to include activity:read_all — the default read scope isn't enough.

Run the setup script (not part of the server itself) to get your first refresh_token, then it lives in strava_tokens.json.

3. Environment variables

Create a .env:

ANTHROPIC_API_KEY=sk-ant-...
STRAVA_CLIENT_ID=...
STRAVA_CLIENT_SECRET=...
STRAVA_TOKEN_PATH=./strava_tokens.json

4. Run it

uv run --env-file .env --with fastmcp fastmcp dev apps mcpserver.py --reload

Project structure

mcpserver.py       # the server: tools, UI, everything
strava_auth.py      # token refresh logic
strava_tokens.json  # your access/refresh token (gitignored)
.env                 # secrets (gitignored)

How the recommendation flow works

1. Dashboard loads → fetches Strava data server-side, renders stats
2. Click "Get Recommendation" → calls get_run_recommendation, a private tool the UI can hit but the model can't call directly in chat
3. That tool re-fetches recent runs, trims to the fields that matter (distance, pace, HR, suffer score), sends them to Claude with a coaching system prompt
4. Response comes back, gets dropped into the page via state

Considerations

• Strava rate limits are tight (100 req/15min, 1000/day)
• max_tokens needs headroom if thinking is ever turned on — thinking and output share the budget