Why Your Webcam Feed Never Leaves Your Browser

When you press play on a cabinet, your browser asks for camera permission, the camera feed starts, and every frame goes to a machine learning model that runs inside the page on your own processor and graphics hardware. The model's output is a short list of numbers: landmark coordinates for your hands, a mesh and expression scores for your face. The game consumes those numbers and nothing else. At no point in that pipeline does a video frame enter a network request. There is no upload step to skip; the architecture simply does not have one.

That is a design choice, not an optimization. A camera game that shipped frames to a server would be laggier, more expensive to run, and would put your living room on infrastructure you cannot see. Running the model in the browser removes the entire category of risk instead of promising to manage it.

It is worth being concrete about how little the game logic touches. Fruit Slice receives fingertip positions and how fast they moved. The Zen Scale receives one angle, the tilt of the line between your eyes. Inferno Drop receives scores like "how open is the mouth" as decimals between zero and one. The camera image itself is drawn to your screen in some cabinets as a backdrop (and several hide it by default and let you toggle it), but the picture is a local rendering, straight from the camera to your own display.

Honesty requires the other half of the list. If you play signed out, the servers see ordinary web traffic: page requests, nothing camera-derived. If you sign in, your account email and the scores you choose to save go to our database, which is what makes the leaderboard work. Analytics and advertising cookies exist only behind the consent banner, and the cookie policy lists them. None of that plumbing has access to the camera pipeline; permission to use your webcam is granted to the page for play, not to any third party. The privacy policy spells all of this out formally.

You do not have to take our word for any of this, and we would rather you did not. Open your browser's developer tools, switch to the Network tab, and play a round. You will see the page load its assets, including the tracking model itself, and then go quiet: no stream of uploads while you wave your arms around, because there is nothing to upload. When you close the tab, the camera indicator light goes off and everything the model computed is gone with the page. Nothing is retained because nothing was recorded.

Webcam games get played by kids, in bedrooms and living rooms, often with other people walking through frame. The only acceptable answer to "where does that footage go" is "nowhere," provable, by construction. It also matters for adoption: people play harder and sillier, which is most of the fun of two-player Air Hockey, when they are not performing for a hypothetical recording. If you are curious about the machinery that makes local processing fast enough for games, that is covered in how browser-based hand tracking actually works.