repMind

A personal Hevy AI training assistant. Connects to your Hevy account, caches your full workout history locally, and gives you:

1. Dashboard — per-lift trend charts (estimated 1RM over time), weekly volume per muscle group, stalled-lift detection, and a coach-voiced "what to improve this week" card.
2. Chat — Claude-powered, reads your real training history and can propose Hevy routines. Nothing is ever pushed to Hevy without an explicit Approve & Push click.

Single-user, self-hosted. Optimized for one person's real data, not multi-tenant scale.

Stack

• Backend: Python / FastAPI + SQLite (SQLModel). Anthropic Python SDK (claude-opus-4-8).
• Frontend: React + Vite + Recharts.
• Hevy: isolated client module (backend/app/hevy/client.py) — all Hevy API quirks live in one file so it's easy to patch as the (early-stage) API changes.

Requirements

• Hevy Pro + a Hevy API key: https://hevy.com/settings?developer
• An Anthropic API key.
• Python 3.11+, Node 18+.

Setup

cp .env.example .env          # fill in HEVY_API_KEY and ANTHROPIC_API_KEY
                              # leave DRY_RUN=true until you've verified the pipeline

# Backend
cd backend
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
uvicorn app.main:app --reload --port 8000

# Frontend (separate terminal)
cd frontend
npm install
npm run dev                   # http://localhost:5173 (proxies /api to :8000)

First run performs a full workout sync on startup. Use the Sync now button (or POST /api/sync) afterward for delta syncs.

Safety model

• HEVY_API_KEY and ANTHROPIC_API_KEY are read server-side only. The frontend never sees them — it talks only to the backend.
• DRY_RUN=true makes every Hevy write log its resolved payload and return a fake id instead of pushing live. Verify the whole flow this way before flipping to false.
• Claude can read your history freely but can only propose routines. A routine is pushed to Hevy only after you click Approve & Push.
• The Hevy API has no DELETE endpoints; repMind builds no delete features.

Coaching context

coach-context.md holds your training philosophy and goals. It's read fresh at the start of each chat session — edit it as your numbers and priorities change; no code edits needed.

Deployment (self-hosted on a tailnet host)

Deployed to pi4host and served at http://pi4host:8000 (tailnet-only — bound to the tailscale IP, not the LAN). Runs as the repmind systemd service (auto-restart, starts on boot); FastAPI serves the built frontend as static files, so the host needs no Node runtime.

Ship updates with one command from this Mac (Tailscale must be up):

./deploy.sh

It builds the frontend locally, pushes main, pulls the backend on the Pi, rsyncs the built UI, reinstalls any new deps, restarts the service, and health-checks. Commit backend changes before running (the Pi deploys backend from pushed main; the frontend ships from the local build). Logs on the host: journalctl -u repmind -f.