Coach AI¶

Your coach, wherever your agent lives.

Coach AI is not an app you run — it's a portable coaching agent definition that installs into agent harnesses you already use (Claude Code and Codex CLI). A small coach CLI configures a coach personality, a set of skills (SKILL.md files), and live MCP connections to your fitness data. The harness's own reasoning, memory, and scheduling (/loop, /schedule) do the rest.

Get started Browse skills

How it works¶

There is no long-running "Coach AI process". At runtime there is just Claude Code or Codex CLI, configured with:

A personality — coaching tone & philosophy, written into CLAUDE.md / AGENTS.md.
Skills — step-by-step SKILL.md procedures for readiness checks, workout generation, training evaluation, plan adjustments, goal research, and personality setup.
MCP servers — live connections to your fitness data (Garmin, Strava, Google/Outlook Calendar).
A local data/ directory — goals, research notes, a thin daily plan, readiness logs, and the rendered coach personality. Nothing duplicates your workout history — your metrics source / calendar is the system of record.

flowchart TD
    A["Choose a path<br/><b>garmin</b> — or —<br/><b>strava</b> + google/outlook calendar"] --> B["coach CLI<br/>setup_source · install"]
    B --> C["Claude Code<br/>CLAUDE.md · .mcp.json · .claude/skills/"]
    B --> D["Codex CLI<br/>AGENTS.md · ~/.codex/config.toml · skills.config"]
    C --> E["Personality · capability-gated skills · MCP servers"]
    D --> E
    E <--> F["Metrics source<br/>(Garmin / Strava)"]
    E <--> G["Workout-calendar source<br/>(Garmin / Google / Outlook)"]
    E <--> H["Local data/<br/>goals · plan · readiness logs"]

Core principle: intelligence lives in the agent, not in Python

The Python in this repo only fetches, normalizes, and persists data (coach/analysis/assemble.py, coach/storage/store.py) — it never computes a fitness score or a "should I train today" verdict. That judgment is the agent's, every time, informed by your goals and the normalized data. See Architecture & principles for the full reasoning.

Supported paths & capabilities¶

Coach AI ships two functional paths. A path = one metrics source + one workout-calendar source (Garmin covers both). The installer resolves the union of your configured sources' capabilities and uses it to gate and tailor every skill — nothing is faked when a capability is missing.

Source	Status	Role(s)	Capabilities
`garmin`	functional	metrics + workout_calendar	`readiness`, `hrv`, `sleep`, `body_battery`, `stress`, `training_load`, `vo2max`, `training_effect`, `structured_workouts`
`strava`	functional	metrics (read-only)	`activity_streams`, `prs`, `training_load`
`google_calendar`	functional	workout_calendar	`free_text_workouts`
`outlook_calendar`	functional	workout_calendar	`free_text_workouts`
`apple_health` / `whoop`	scaffold	metrics	future — no MCP yet

The two paths are not equal, and the coach says so up front:

Garmin — full readiness/HRV/sleep/body-battery/training-load data, plus structured workouts pushed to the Garmin calendar and rescheduled in place.
Strava + Google/Outlook Calendar — Strava supplies activity streams and PRs; the calendar is the workout system-of-record, with each session written as a free-text event (description/body). No readiness/HRV/ sleep/body-battery — those steps are skipped, not faked.

See Capabilities & paths for the full capability matrix.

Where to go next¶

Installation

Authenticate a source, run coach setup, install into your harness, and set up the daily loop.

Installation guide
Architecture & principles

The five design principles, the high-level architecture, and the harness installer contract.

Concepts
Skills catalog

What each skill does, when it runs, and how it behaves on each path.

Skills
Daily workflow

Onboarding, the daily loop, and annotated example conversations.

Daily workflow