Camera. The page asks for your selfie camera with getUserMedia({ video: { facingMode: 'user' } }). That only works in a secure context (HTTPS or localhost). Frames stay in the page — nothing is uploaded.
Tracking. Open-source MediaPipe models run in WebAssembly/WebGL inside your browser. Face mode uses the Face Landmarker (478 landmarks + 52 ARKit-style blendshapes — the browInnerUp/browDown shapes are proxies for your frontalis and corrugator muscles). Body mode uses the Pose Landmarker (33 keypoints).
Calibration. Everyone's resting face and range of motion is different, so you first show the game your neutral pose and your extremes. Thresholds are set from your range (about 45% of it) with a noise floor, and the neutral baseline keeps drifting slowly while you play.
Signal processing. Continuous signals (head tilt, lean) go through a 1€ filter — low jitter when still, low lag when moving — then a deadzone. Discrete actions (jump, duck, boost) use a Schmitt trigger: separate on/off thresholds (hysteresis) plus a refractory period, so a shaky signal can't fire twice.
Performance. Inference is paced to stay under about half of each frame's budget and the camera image is downscaled in steps (640 → 256 px) when the phone slows down or throttles from heat. Rendering drops device-pixel-ratio the same way. Everything pauses when you leave the frame or switch tabs.
Ergonomics. Face mode needs no raised arms, so there's no "gorilla arm". Body mode uses quick gestures instead of held poses, and every ~40 seconds there's a short breather with no obstacles.
Keys / touch. ← → or A/D steer, ↑ / W / Space jump, ↓ / S duck, Shift boost. On touch: drag sideways to steer, flick up to jump, flick down to duck.