# Relay

A fully client-side, real-time audio translator built on Gemini Live. Relay captures audio from a user-selected tab, window, or screen and streams 16 kHz PCM directly from the browser to `gemini-3.5-live-translate-preview`. Translated 24 kHz PCM audio and transcripts are rendered in the browser.

## Run locally

```bash
npm install
npm run dev
```

Open the localhost URL, enter a Gemini API key, choose a target language, and click **Share audio & translate**. In the browser picker, select a source and explicitly enable audio sharing.

## Browser requirements

- A secure context (`https://` or localhost)
- `navigator.mediaDevices.getDisplayMedia()` with audio capture support
- `AudioWorklet` and Web Audio
- Chrome or Edge is recommended for tab/system audio sharing
- Output-device selection appears only when `HTMLMediaElement.setSinkId()` is supported

Prefer sharing the individual tab that is playing the source audio. Full system-audio capture may also include Relay's translated output in the digital mix. Relay requests `selfBrowserSurface: "exclude"` and `restrictOwnAudio: true` where supported; a separate output device can also help when the operating system isolates its audio from the captured source.

## Privacy and key handling

There is no application backend. The WebSocket connects directly to Google’s Gemini Live endpoint, and the Gemini API key is stored in `localStorage` under `relay-gemini-api-key`. This bring-your-own-key design means any script or browser extension with access to this origin may be able to read the key; deploy the SPA only from a trusted origin.

## Checks

```bash
npm run build
npm run lint
```