coach-mcp

An opinionated AI training coach as an MCP server. It pulls your real data from Garmin Connect and prescribes with authority — science-based load management (ACWR), code-enforced injury gates (the server rejects plans that violate an active injury restriction, no matter what the LLM says), and persistent coaching memory so decisions, rationale, and your adaptation patterns survive between conversations. It will tell you "no" when your enthusiasm exceeds your capacity.

All health data and credentials stay on your machine — see Security & Privacy.

Quickstart

You need Python 3.12+, a free Garmin Connect account, and an MCP client (Claude Code, Claude Desktop, or Cursor).

Option A: uvx (recommended)

No install step — your MCP client runs the server on demand:

uvx garmin-coach-mcp

Jump to Connect your MCP client and use uvx as the command.

Option B: from source

git clone https://github.com/snoozelieb/coach-mcp.git
cd coach-mcp

python -m venv .venv
# Linux/macOS:
source .venv/bin/activate
# Windows:
.venv\Scripts\activate

pip install -r requirements.txt
cp .env.example .env   # then edit: GARMIN_EMAIL, GARMIN_PASSWORD
python server.py

Connect your MCP client

The server needs two environment variables: GARMIN_EMAIL and GARMIN_PASSWORD. Optional: COACH_DATA_DIR (where your coaching data lives) and ANTHROPIC_API_KEY (only for the standalone daily_loop.py --llm script). From a source checkout, a .env file works too.

Claude Code

claude mcp add coach-mcp \
  --env GARMIN_EMAIL=you@example.com \
  --env GARMIN_PASSWORD=your_garmin_password \
  -- uvx garmin-coach-mcp

Or in .mcp.json:

{
  "mcpServers": {
    "coach-mcp": {
      "command": "uvx",
      "args": ["garmin-coach-mcp"],
      "env": {
        "GARMIN_EMAIL": "you@example.com",
        "GARMIN_PASSWORD": "your_garmin_password",
        "COACH_DATA_DIR": "/path/to/your/coach-data"
      }
    }
  }
}

Running from source instead: claude mcp add coach-mcp -- python /full/path/to/coach-mcp/server.py

Claude Desktop

In claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "coach-mcp": {
      "command": "uvx",
      "args": ["garmin-coach-mcp"],
      "env": {
        "GARMIN_EMAIL": "you@example.com",
        "GARMIN_PASSWORD": "your_garmin_password",
        "COACH_DATA_DIR": "/path/to/your/coach-data"
      }
    }
  }
}

Cursor

In .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "coach-mcp": {
      "command": "uvx",
      "args": ["garmin-coach-mcp"],
      "env": {
        "GARMIN_EMAIL": "you@example.com",
        "GARMIN_PASSWORD": "your_garmin_password",
        "COACH_DATA_DIR": "/path/to/your/coach-data"
      }
    }
  }
}

If you installed with pip install garmin-coach-mcp instead of uvx, use "command": "garmin-coach-mcp" with no args in any of the blocks above.

First run

1. Create your profile. From a source checkout, run the interactive wizard:

   bash

   python scripts/setup_wizard.py
   

   It creates your athlete profile, training config, and empty plan/memory files in the data directory. Alternatively, create the two required files by hand and let the coach fill in the rest via conversation:

   bash

   echo '{"personal":{"name":null},"injury_history":[],"life_constraints":{}}' > data/athlete.json
   echo '{"events":[],"current_block":{"phase":"base"}}' > data/training_config.json
   

2. Pull your Garmin baseline. In your MCP client, say:

   "Run refresh_athlete_baseline and set up my training."

   The coach pulls your name, weight, age, HR data, and training capacity from Garmin, then starts the onboarding conversation — goals, constraints, injury history, race calendar.

3. Garmin MFA / expired session. Garmin logins are token-cached. If tools start returning AUTH_REQUIRED, recover with:

   bash

   python scripts/garmin_login.py
   

   It does a fresh credential login, prompts for the MFA code if Garmin asks, and saves new tokens. Restart the MCP server afterwards.

How it works

1. Snapshot first — every coaching conversation starts from get_coaching_snapshot(): current time context, 7-day week grid (rest days explicit), fitness metrics, plan adherence, open anomalies, injuries, sleep gate.
2. Load hierarchy before prescribing — overall ACWR (injury gate, 0.8–1.3 sweet spot), then sport-specific ACWR (spike detection), then sport-specific CTL (race readiness).
3. Hard gates are code, not vibes — update_weekly_plan and push_plan_to_garmin reject sessions that violate an active injury's restricted activities, and every non-rest session must carry a purpose or the save is refused.
4. Curiosity with memory — planned-vs-actual anomalies (missed session, type mismatch, activity on a rest day) register once with a lifecycle (open → asked → resolved); the coach asks you what happened instead of silently assuming.
5. Everything persists — decisions, approvals, adaptation patterns, and season lifecycle (race debriefs, phase transitions) live in local JSON and carry across sessions.

MCP surface

49 tools — you don't call them directly; the coach uses them during conversation:

Every tool carries MCP annotations (read-only / destructive / idempotent / open-world), enforced by tests.

5 prompts: weekly_planning, morning_brief, injury_assessment, week_review, onboarding.

6 resources: coach://athlete/profile, coach://plan/current, coach://config/training, coach://coaching/decisions, coach://context/now, coach://coaching/doctrine (the long-form coaching doctrine).

Security & Privacy

Everything stays on your machine:

• Credentials: GARMIN_EMAIL/GARMIN_PASSWORD live in your MCP client config or a local .env. Garmin OAuth tokens are cached in a local token store (.garth/ in a source checkout, a per-user garmin-tokens directory for installed copies; override with COACH_TOKEN_DIR).
• Health data: all coaching data (profile, plans, fitness history, sleep, coaching memory) is local JSON in your data directory. There is no backend, no telemetry, no analytics.
• What leaves your machine: requests to Garmin's own API (your credentials/tokens, sent only to Garmin); whatever your MCP client sends to its LLM as part of the conversation; optional public web-page fetches when the coach researches a race, injury, or exercise; and, only if you run daily_loop.py --llm, one request to the Anthropic API.
• Single athlete per data directory by design. For multiple athletes, run separate server instances with separate COACH_DATA_DIRs.

See SECURITY.md for details and how to report issues.

Data directory

Resolution order: COACH_DATA_DIR env var → data/ in a source checkout → a per-user data directory (created on first run for installed packages). The only file shipped with the package is methodology.json (safety rules, race templates, personas); everything personal is created locally and never committed.

Advanced

# HTTP transport (streamable-http) instead of stdio
COACH_TRANSPORT=streamable-http FASTMCP_PORT=8000 garmin-coach-mcp

# Code Mode (search/execute meta-tools instead of 49 individual tools)
pip install fastmcp[code-mode]
COACH_CODE_MODE=1 garmin-coach-mcp

# Standalone morning audit
python scripts/daily_loop.py          # template-based brief
python scripts/daily_loop.py --llm    # LLM brief (needs ANTHROPIC_API_KEY)

# Tests (1,333 tests; clean checkouts use committed sanitized fixtures)
pip install -r requirements-dev.txt
python -m pytest -q

Architecture

server.py registers tools from the coach/ package (11 tool modules, pure parsers, a typed pydantic storage layer, CTL/ATL/ACWR fitness math, a Garmin client with token-first auth, and a workout builder that pushes structured workouts to your watch). The project went through a five-phase modernization — auth rebuild, schema layer, hard gates, sectioned snapshot, packaging — whose full history and rationale live in docs/UPGRADE_ROADMAP.md. Development conventions are in CLAUDE.md.

License

MIT