This measures your heart rate from a webcam, with nothing touching you. Blood flowing through the skin of your face changes its colour by a fraction of a percent with every beat - far too little to see, but a camera can pick it out if you sit still in decent light.
Face tracking places sampling patches on your forehead and both cheeks, skipping hair, beard and glasses. The red, green and blue channels are combined so that changes caused by movement and lighting cancel out and the pulse survives. Fifteen seconds of that signal are then examined for the frequency that stands out.
Pulse rate only. Heart rate variability, rhythm and atrial fibrillation screening, and pulse transit time were all built and then removed: at 30 frames per second the timing of individual beats is too coarse to support them, and the earlier versions produced confident-looking numbers that were wrong.
Movement, dim or uneven light, backlighting, and talking. It refuses to show a number rather than show a bad one, so a blank reading is normal and means the estimate was rejected. Expect to spend the first few seconds adjusting your position and light.
Checked against a Pixel Watch ECG recorded within a minute of each session. When the signal quality indicator is above its threshold the estimate landed within about 1 bpm of the ECG. When quality was marginal it was out by 3 bpm, while still looking perfectly confident on screen. That is from three sessions with one person - nowhere near enough to claim a general accuracy figure, and nothing here has been tested across different skin tones, ages, lighting setups or cameras.
Video is processed entirely in your browser. No images and no measurements are uploaded or stored anywhere. Nothing leaves the page unless you press Log Data, which saves a file to your own device.
Press Start Camera and allow camera access. Sit still, face a light source and avoid
backlighting. The estimate needs a full 15 second window and only appears while signal quality clears 50%,
so a blank reading means it was rejected as untrustworthy - not that nothing was detected.
Read the bar across the top of the video rather than this panel while measuring. Looking
away moves your head, and head movement is the single biggest cause of a failed reading.
Stillness on that bar is the number you can act on: on test recordings every window that
stayed above roughly 60% cleared the quality gate, and everything below 45% failed.