This is Part 4 of a series on building an agentic personal trainer. Read Part 1 for architecture, Part 2 for tools, and Part 3 for the system prompt.

A personal trainer who doesn't know your recent workouts isn't very personal. The agent needed to connect to Garmin Connect—where my watch, bike computer, and other services like Wahoo SYSTM sync workout data automatically.

The garmin-connect npm package handles authentication. On every chat startup, the system syncs the last 14 days (garmin.js:210-235):

const client = new GarminConnect({
  username: process.env.GARMIN_EMAIL,
  password: process.env.GARMIN_PASSWORD
});

await client.login();
const activities = await client.getActivities(0, 14);

*Garmin sync flow: Garmin activities flow from watch to API to normalized SQLite records*

Garmin's activity types are granular—lap_swimming, road_biking, virtual_ride. This required a mapping function to normalize them (garmin.js:18-48):

function mapGarminActivityType(garminType) {
  const mapping = {
    'running': 'run',
    'cycling': 'bike',
    'road_biking': 'bike',
    'virtual_ride': 'bike',
    'lap_swimming': 'swim',
    'open_water_swimming': 'swim',
    'strength_training': 'strength',
    'yoga': 'yoga',
    'walking': 'walk',
    'hiking': 'hike'
  };
  // Pattern matching for partial matches
  for (const [key, value] of Object.entries(mapping)) {
    if (typeStr.includes(key)) return value;
  }
  return 'other';
}

Each activity comes with duration (in seconds), distance (in meters), heart rate, and calories. The system converts to minutes and kilometers, then builds a notes string with the useful metadata (garmin.js:136-152). This is a prototype simplification—pool swims are typically measured in yards or meters, not kilometers, which would be a good enhancement for a production system:

const notes = [];
if (activity.activityName) notes.push(activity.activityName);
if (activity.averageHR) notes.push(`Avg HR: ${activity.averageHR}`);
if (activity.maxHR) notes.push(`Max HR: ${activity.maxHR}`);
if (activity.calories) notes.push(`${activity.calories} cal`);

The sync runs silently at chat startup. If Garmin is unreachable (rate limits, auth issues), it gracefully falls back to existing data with a warning. The agent doesn't need to know the sync happened—it just calls get_recent_workouts and gets fresh data.

Auto-sync on startup was the right call for this prototype. Manual sync commands add friction. The agent should just know what you've been doing without you having to tell it. Refinement to this sync process is another good area of improvement for this prototype.

But it surfaced a new problem: I use both a Garmin watch and Wahoo SYSTM for indoor training. Same workout, two records. In some cases, the agent thinks I'm training twice as much as I am.

That duplicate detection problem is next.

Part 4 of 9 in the Agentic Personal Trainer series.

agentic-personal-trainer - View on GitHub

This post is part of my daily AI journey blog at Mosaic Mesh AI. Building in public, learning in public, sharing the messy middle of AI development.

Building an Agentic Personal Trainer - Part 4: Garmin Integration

Building an Agentic Personal Trainer - Part 5: Smart Duplicate Detection

Building an Agentic Personal Trainer - Part 3: The System Prompt

Mosaic Mesh AI