Voice Browse

Voice Browse is a free revolutionary Chrome extension that transforms your web browser into a completely hands-free platform controlled entirely by your voice. Designed with zero configuration in mind, it allows you to open websites, search the web, scroll, type, and take screenshots using natural speech in English, Bahasa Melayu, or 中文 (Mandarin)—even supporting seamless language switching mid-sentence. The tool handles complex web layouts effortlessly by overlaying numbered hints onto dense, clickable elements, allowing you to simply call out a number to select it. If a command is misunderstood, smart corrective phrases like "no, not that one" or "undo" instantly reverse or redirect the action, making it an ideal accessibility aid for anyone with motor impairments, RSI, or a preference for voice control. At the core of Voice Browse is a highly flexible dual-AI engine that lets you choose between the decision engine such as web models (kev,Laya) model or ollama offline model(tev1) or openai decision api luna or the online TypeSafe Jev endpoint based on your specific needs. The built-in Laya model is completely free, runs 100% offline, and utilizes WebGPU technology to process commands instantly on your computer's graphics card without needing an account or API key. Because all data remains strictly on your device, Laya offers maximum privacy for daily browsing. For power users navigating exceptionally intricate websites, Voice Browse optionally connects to TypeSafe Jev. This online endpoint acts as a premium upgrade, using cloud-based processing via your own API key to deliver state-of-the-art decision accuracy on the most challenging user interfaces. To get started, simply install the extension on Chrome 113 or newer, click the toolbar icon, and grant microphone access to begin browsing the web entirely by voice. Permission justifications (Privacy practices tab) storage — Settings are saved locally in chrome.storage.local (decision endpoint, voice language, saved tab groups, theme, on/off state). Nothing is synced; storage persists your preferences between sessions. tabs — Reads the active tab's URL/title so commands can act on the right page ("go back", "search this site"), and switches, opens, closes, and manages tabs by voice. downloads — Saves screenshots taken by the "print screen" / "tangkap skrin" voice command to your Downloads folder. bookmarks — Adds and removes bookmarks by voice ("bookmark this page" / "remove bookmark") and opens the bookmarks manager ("open my bookmarks"). offscreen — A hidden offscreen document hosts the free offline decision model (ONNX Runtime) and the microphone (Web Speech API), so voice commands keep working across tabs and windows instead of only while the toolbar popup is open. host permission (<all_urls>) — Runs the content script on web pages so voice commands can click, type, and scroll on them; downloads the free offline model weights from Hugging Face; and optionally reaches the decision endpoint you configure (e.g. a localhost proxy). It never reads or uploads your browsing history. Remote code justification No remote code is executed. Manifest V3 requires all JavaScript to be bundled and reviewed: ONNX Runtime Web and the tokenizer are vendored in the package. The only network content downloaded is model data — weights for the free offline decision model (~0.8–1.3 GB, cached by the browser) — which is interpreted by the bundled, reviewed ONNX Runtime, not executed as code. Optional online mode sends your current command context to the decision endpoint you explicitly configure. Data usage certification (Privacy practices tab) Answer the data-safety form as: no personal, health, financial, authentication, or location data; no audio is ever transmitted. In the optional online mode only, the current spoken command and a compact snapshot of the active page are sent to the decision endpoint you configure — solely to pick the action. Offline mode sends nothing. Then tick "I certify that my data usage complies with our Developer Program Policies."