Try the app — RefereAI MLX

Path A · Quickstart (recommended)

One command. Five minutes. Live.

The quickstart script installs MLX, fetches YOLO26 weights, runs a smoke test, and leaves you at a prompt ready to refereai serve. Apple Silicon Mac, Python 3.10–3.12, and a phone on the same Wi-Fi — that's the whole prerequisite list.

# Quickstart — clone + bootstrap in one shot
curl -LsSf https://raw.githubusercontent.com/jravinder/refereai-mlx/main/scripts/quickstart.sh | bash

# Then start the live server (prints a QR code with the LAN URL)
refereai serve

5 minutes from a fresh shell to a live overlay on a phone.
Apple Silicon only. M1, M2, M3, M4 — any tier. Intel Macs are not supported.
Then refereai serve and scan the QR printed in your terminal from any phone on the same Wi-Fi.

View source on GitHub → Watch demo videos

Path B · Manual install for developers

Or: clone, bootstrap, inspect every step.

If you want to read the bootstrap before running it (which: good), the long-form path is clone + script. Same outcome, more visibility.

# Manual — clone first, then bootstrap
git clone https://github.com/jravinder/refereai-mlx && cd refereai-mlx
bash scripts/bootstrap.sh

# Inspect, run tests, then serve
.venv/bin/python -m pytest tests/ -v
refereai serve

Read the setup docs →

What to expect when it works

Three beats. No cloud in any of them.

A reasonable session, end-to-end, on a kitchen-table M2 with a phone on the same Wi-Fi.

Scan the QR. Terminal prints a QR with the LAN URL. Your phone browser opens directly to the PWA. 8 seconds.

Point at a sport. Classifier locks on. Boxes + rationale render on the phone, no perceptible lag. Instant.

Verify it's local. nettop shows zero WAN bytes. Flip airplane mode on — the overlay keeps rendering. No cloud.

Hardware support

It works on these Macs.

Anything M-series. The bootstrap script picks the right MLX wheel for your chip — you don't have to think about it.

M1 M1 Pro M1 Max M1 Ultra M2 M2 Pro M2 Max M2 Ultra M3 M3 Pro M3 Max M4 M4 Pro M4 Max

Intel Macs are not supported — yolo26-mlx requires Apple Silicon. The unified-memory inference path is the whole reason this architecture works.

Honest scope

It does NOT work on these (yet).

Telling you what doesn't work is how you can trust what does. None of the below are on the roadmap for v1.

Intel Macs. No x86 inference path. MLX is Apple Silicon only.
Windows or Linux. The whole product is built around Apple's unified memory + Metal stack. A CUDA fork exists in the author's other repos but is not this project.
Macs with less than 16 GB RAM. The smoke test passes on 8 GB, but the end-to-end pipeline plus a browser plus Safari Web Inspector will swap. 16 GB is the floor for a good demo.
Macs without Xcode Command Line Tools. xcode-select --install first — the bootstrap script will detect and stop early if it's missing.
Without a phone on the same Wi-Fi. No LAN, no QR scan. The PWA is the camera; there is no Mac-webcam fallback in v1.

Why hosted debug URLs 404

The camera and HIL Review console live on your Mac.

The hosted domain is the public landing site. The capture app and HIL Review pages are served by the live Python process at refereai-mlx.local, your LAN IP, or Tailscale on port 8765. That is why refereai.xyz/debug/videos/review is a Vercel 404, while the local Mac URL opens the protected HIL Review console.

After refereai serve, open the local URL printed in your terminal. Debug pages are password protected when REFEREAI_DEBUG_PASSWORD is set.

See it running → Who it's for