Loading page...Загружаем страницу...正在加载页面...
AtlasRepo

I built a self-hosted AI running coach that reads my Garmin and plans my runs and lifts

Description and use

The author built a self-hosted AI running coach that reads Garmin data, plans runs and lifts, saves an hour per week, and helped achieve a sub-20 5K goal.

Problem

The author needed a more efficient way to plan and evaluate running and lifting sessions without manually uploading Garmin screenshots to Claude.

Outcome

The author achieved a sub-20 5K goal within a month, reduced planning time by an hour per week, and gained a more data-driven, adaptive training plan.

Workflow

  • Set up a self-hosted Hermes agent on a Mac mini
  • Configure Garmin read access via MCP, Strava uploads, and a notes folder
  • Write a detailed instruction file describing coaching preferences
  • Agent sends a daily 8am message, checks sleep and HRV, asks how the user feels
  • Agent designs a run plan and writes it to Garmin Connect
  • Agent polls Strava for uploaded runs, pulls laps, grades the session against the prescribed plan
  • Agent tracks times, distances, lift numbers, and adjusts load progression
  • Agent provides nutrition guidance around hard sessions
  • Agent adapts plans based on recent performance and load

Evidence

I set a sub-20 5K goal and got there about a month later.

Source

It occasionally reads Garmin's readiness score too literally, so I told it to ignore that metric.

Source

  • Record date: 2026-09-18T23:45:39.375Z