coachd
A self-hosted gateway between your AI coach and your Garmin watch. It serves your Garmin data to any AI agent and puts the workout you agreed on onto your watch, but only after you approve it.
One signed Docker image, one instance per runner. Your Garmin key and health data stay on your server (data and privacy).
📖 Documentation: stefiix92.github.io/coachd: install, connect an agent, watch app, MCP tools, FAQ, troubleshooting.
Status: early and built by one runner for their own training, but it works end to end: Garmin data in, MCP out, passkey-approved workouts to the watch. Expect rough edges. While the version is 0.x, a minor release may change configuration or behaviour, so read the changelog before upgrading.
Why
- Talking training through with an AI (reviews, how you feel, weather, niggles) works better than a black-box plan like Garmin's Daily Suggested Workouts.
- A plain AI chat can't see your watch data, forgets context between chats, and can't put a workout on your watch.
- An AI must never push anything to your watch on its own.
coachd is a gateway, not a coach: the AI agent and you decide the training; coachd provides the data, keeps the context, warns, and delivers only after your explicit approval.
What it does
flowchart LR
garmin["Garmin Connect<br/>Personal API"] -->|"activities, FIT files,<br/>sleep, HRV, ..."| coachd[("coachd")]
agent["AI agent<br/>(Claude, ChatGPT, ...)"] <-->|"MCP: data,<br/>goal, notes, proposals"| coachd
coachd -->|"ntfy: workout ready"| phone["You, on the<br/>approval page"]
phone -->|"approve with passkey"| coachd
coachd -->|"approved workouts"| watch["Watch app,<br/>FIT or Zwift file"]
- Data in. Activities with their original FIT files, sleep, HRV, resting HR, Body Battery and more, from the official Garmin Connect Personal API, plus simple totals. Numbers, never verdicts.
- Context. The goal the agent sets and a few durable notes, so any agent or new chat starts informed.
- Workout out. The agent proposes a structured workout; you approve it with a passkey; coachd delivers it through the Connect IQ watch app, as a FIT file, or as a Zwift workout.
- Your own rating. Your post-workout "How did you feel?" and RPE reach the agent too.
- Planned vs actual. The executed steps of each workout, linked to the workout coachd delivered.
There is no chat UI and no built-in LLM. You talk to Claude (or ChatGPT, Gemini, Cursor, Open WebUI or a local model), and the agent calls coachd over MCP. No MCP tool can approve, deliver or withdraw a workout: only you can, with your passkey.
Read how it works and an example week, then install it. You need your own Garmin Connect Personal API key and a Docker host. See also compatibility and how coachd compares.
Stack
Go, the official MCP Go SDK, SQLite (pure Go), muktihari/fit, go-webauthn, and server-rendered HTML with no npm. The watch app is Connect IQ (Monkey C). Packaging is a signed multi-arch image plus deploy/compose.yaml with Tailscale, Grafana Alloy, VictoriaMetrics and VictoriaLogs.
Design records are in docs/design/, starting with coachd-design.md. The documentation site's source is in docs/content/.
Roadmap
Next: a smoother first-time install, and wider live testing of AI clients against built-in mode. Planned work is tracked in roadmap issues; smaller ones are labelled good first issue. Changes are listed in CHANGELOG.md.
Disclaimer
Not affiliated with or endorsed by Garmin. Not medical advice. If something hurts, stop and see a professional.
Contributing and security
- Contributions are welcome: see
CONTRIBUTING.md. Commits need a DCO sign-off.
- Found a vulnerability? Report it privately, as described in
SECURITY.md. Please don't open a public issue.
Support
coachd is free and built in spare time. If it helps your training, you can buy me a coffee on Ko-fi.
License
GNU Affero General Public License v3.0 (AGPL-3.0-only). You may use, modify and self-host coachd freely. If you run a modified version as a service for others, you must publish your changes under the same license.
