I’m writing this from a squat rack.

Not metaphorically. I just finished five sets of deadlifts at 280, and my AI coach — an agent named Zephyr running on a Mac mini in my house — tracked every set, called my rest periods, told me when to drop weight, and pulled my recovery data from the Whoop API before I even started warming up.

No app. No subscription fitness platform. Just an agent, an API, and voice memos.

How It Works

The setup is simpler than you’d think:

1. Whoop API pulls my recovery score, HRV, resting heart rate, and SpO2 before the session. Green light means we push. Red means we back off.

2. Voice memos are the interface. I talk to my phone between sets — “three reps at 285, heavy” — and the agent transcribes it locally, logs the data, and responds with coaching cues.

3. Progressive overload logic lives in the agent’s memory. It knows my last session’s numbers, my program structure (5×5 strength core, hypertrophy accessories), and the rules: hit all 25 reps clean before adding weight. Two consecutive failures at a weight means deload 10%.

4. Session history persists across workouts. The agent doesn’t just know today — it knows last week, last month, and the trend.

What Happened Today

Recovery was 87% — green across the board. The agent bumped my deadlift from 280 to 285. I hit 3 reps on set 1 (cold), ground out 5 on set 2, fell back to 3 on set 3. The agent called it: drop to 280, finish clean, prescribe 275 for next session.

That’s not a generic “great job!” from an app. That’s a real coaching decision based on my data, my history, in real time.

Why This Matters

Every fitness app wants to be your coach. They’ll generate a workout plan from a questionnaire and send you push notifications. But coaching isn’t a template — it’s adaptation. It’s the decision to drop weight mid-session because your rest periods aren’t long enough at near-max loads. It’s knowing that you missed creatine yesterday and factoring that into expectations.

The stack is:

• Whoop for biometrics (recovery, strain, sleep)
• OpenClaw for the agent runtime
• Local transcription (mlx-whisper on Apple Silicon — nothing leaves the device)
• Discord as the chat interface (voice memos in, coaching cues out)

Total monthly cost beyond the Whoop membership: roughly what you’d pay for one protein shake.

The Bigger Picture

This is what personal AI agents actually look like. Not chatbots answering trivia. Not copilots writing code. An agent that knows your body, your program, your goals — and makes real decisions with you in real time.

I build these for a living at wade.digital. If you want an agent that actually does something — coaching, business ops, content, whatever your domain is — that’s what we make.

Now if you’ll excuse me, I have squats to do.