WebLLM

Client-side screen-aware assistant powered by a local vision-language model.

WebLLM — private screen-aware AI that runs inside Chrome WebLLM is a Chrome extension that lets you ask questions about what’s on your screen — or chat in plain text mode — using a vision-language model that runs on your computer inside the browser. Your questions are not sent to a WebLLM server. There is no analytics, no account system, and no saved chat history. If you’ve wanted AI help while browsing but don’t want to install heavy desktop AI software, or your PC isn’t built for running local models the traditional way (large downloads, command-line tools, GPU-only setups), WebLLM is built for you. Why install WebLLM? 1. For people whose computer “can’t run local AI” the usual way Many “local AI” tools expect a powerful GPU, lots of RAM, and separate apps like Ollama, LM Studio, or Python environments. That leaves out a lot of normal laptops and work machines. WebLLM takes a different approach: It runs inside Google Chrome, not as a separate AI server on your PC It uses WebGPU when available (and falls back to CPU if not) There is nothing to configure — install the extension, wait for the one-time model download, and start asking questions Screen OFF mode lets you chat without screenshots, which is lighter and works in more situations You don’t need to be a developer. You don’t need CUDA drivers or a terminal. If Chrome runs on your machine, WebLLM can run there too. 2. It runs in the browser — simple and self-contained WebLLM is designed around how you already work: Click the extension icon → ask a question → get a streaming answer Toggle Screen on to capture the visible tab and ask about what you see Toggle Screen off for text-only questions (no screenshot) One request at a time — use Stop if you want to cancel First launch downloads the model from Hugging Face once (~3 GB). After that, Chrome can reuse cached weights so later sessions start much faster. Day-to-day use does not require constant re-downloading. 3. For people who value safety and privacy WebLLM is built for users who care about where their data goes: No WebLLM backend — inference happens locally in the extension runtime No telemetry or analytics No saved chat history — prompts and screenshots are cleared after each request Screen capture only when you turn Screen on and send a message — nothing is captured in the background Network use is limited to downloading public model files from Hugging Face (first setup / cache refresh), not sending your questions to a proprietary cloud Your browsing context stays on your device for the duration of a request, then memory is cleared. Model weights may be cached in browser storage so you don’t re-download ~3 GB every time — that cache is for the model only, not your conversations. If privacy, control, and transparency matter to you, WebLLM is a straightforward choice: local processing, no account, no chat log. What can you use it for? “What does this button do on this page?” “Summarize what I’m looking at.” “Explain this error message on screen.” “What are the main options on this settings page?” General questions in Screen OFF mode without capturing anything WebLLM is a helper for understanding what’s in front of you, not a replacement for professional advice (medical, legal, financial, etc.). How it works: You type a question in the popup. If Screen is on, the extension captures a compressed image of your active tab (normal websites — not Chrome internal pages like chrome://). A local Gemma vision model runs in an offscreen Chrome context. The answer streams into the chat. When finished, request data is discarded. Requirements Google Chrome 124 or newer WebGPU recommended for best speed (CPU fallback is supported but slower) ~4 GB free disk space for the one-time model download Internet for the initial model download (offline use possible after the model is cached) Permissions (why we ask) Active tab / tabs — to capture the visible page when Screen mode is on Offscreen — to run the model in a stable Chrome context Hugging Face — to download the public model weights We ask for what the product needs and nothing more. Open source WebLLM is open source. You can review the code, build it yourself, and see exactly how it behaves: https://github.com/yelloworangebananaa/WebLLM A note on expectations WebLLM runs a real vision-language model in the browser. That’s powerful, but: First launch requires a large one-time download CPU-only machines will be slower than WebGPU Answers can be wrong — always verify important information Screen mode does not work on Chrome’s internal pages (chrome://, extension pages) Install WebLLM if you want AI help while browsing — without sending your screen and questions to a company’s cloud, without installing a heavy local AI stack, and without leaving a chat history behind.