Skip to content

README.md

Latest commit

 

History

History
108 lines (87 loc) · 4.73 KB

README.md

File metadata and controls

108 lines (87 loc) · 4.73 KB

fit-agent

Turn your AI agent into your personal fitness coach.

fit-agent is a Go CLI that hands your training data to an AI agent in a form it can actually reason about — and lets the agent push tomorrow's workout straight to your watch or bike computer.

It bridges intervals.icu (where your devices already sync) and an OpenClaw workspace (where your agent does its thinking).

Status: pre-alpha. Design is set, code is being written. See agent-plan.md for the v1 roadmap.

What it does

  • Pulls everything you do — activities, per-lap and per-interval metrics from the underlying .fit file, daily wellness (HRV, RHR, sleep, steps, stress) — from intervals.icu into a local workspace.
  • Renders it for an LLM — concise, commented YAML for activity and wellness data; markdown for narrative files. No 174-field JSON dumps, no token-eating prose tables.
  • Keeps the originals — raw icu JSON and raw .fit files live in .cache/, so the agent-facing files can always be regenerated.
  • Lets the agent plan and push — bundled coaching skills teach an agent to interview you about goals, pick a methodology (Jack Daniels, 80/20, Norwegian, etc.), build a multi-week plan, translate it into daily workouts, and push them to intervals.icu (and on to your device) with one command.
  • Stays out of your way — agent-owned files (fit-agent/ATHLETE-PROFILE.md, fit-agent/TRAINING-PLAN.md, planned-workout intent) are never overwritten. Machine-owned data files are regenerated on every fetch.

Workspace at a glance

my-coaching/
├── skills/                       # OpenClaw coaching skills (bundled)
│   ├── training-plan-coach/
│   ├── workout-builder/
│   └── training-session-coach/
└── fit-agent/
    ├── ATHLETE-PROFILE.md              # goals, history, constraints (you + agent)
    ├── TRAINING-PLAN.md                # the plan the coach builds with you
    ├── activities/2026-05-03.yaml      # today's session(s), with laps
    ├── wellness/2026-05.yaml           # the month's daily wellness
    ├── planned-workouts/2026-05-04.md         # agent-authored workout
    ├── planned-workouts/2026-05-06.4711.icu.md # read-only mirror of an icu-side workout
    └── .cache/                          # raw icu JSON + .fit files

Commands (planned for v1)

fit-agent init                    # one-time setup; scaffolds the workspace
fit-agent fetch --since 30d       # pull activities + wellness + planned
fit-agent sync-workouts           # push agent-authored workouts and pull icu-side ones
fit-agent serve                   # poll intervals.icu on a cadence (daemon)
fit-agent setup-service           # install ~/.config/systemd/user/fit-agent.service
fit-agent remove-service          # tear it down again

Plus atomic subcommands for debugging and agent use:

fit-agent cache activity <id>     # download raw json + .fit only
fit-agent render activity <id>    # cache → YAML, no network
fit-agent fit laps <file.fit>     # inspect a parsed .fit file
fit-agent workout render <file>   # convert the fit-workout DSL

How it works with your agent

  1. fit-agent init creates the workspace and installs three coaching skills as OpenClaw workspace skills.
  2. You open the workspace with your agent. The agent reads ATHLETE-PROFILE.md and recent activities/ + wellness/ data.
  3. The training-plan coach skill walks you through goals and methodology, then writes TRAINING-PLAN.md.
  4. The workout-builder skill turns the plan into concrete daily workouts under planned-workouts/ and runs fit-agent sync-workouts, which pushes new files to intervals.icu and pulls back any workouts authored on icu (or by another device) as read-only .icu.md mirrors.
  5. The training-session coach skill checks today's wellness against today's planned workout and recommends adjustments before you train.

The CLI never calls an LLM. The agent calls the CLI.

Roadmap

  • v1 — three commands, intervals.icu only, manual fetch.
  • post-v1 — fit-agent serve polls intervals.icu on a cadence (15 min default, quiet-hours aware), wrapped by setup-service / remove-service for systemd-user installation. OpenClaw webhook integration so the agent is notified the moment new data lands, and full coaching prompts in the bundled skills.

Built on

License

MIT — see LICENSE.