A browser extension that uses diffusion LLMs to help you navigate and simplify any webpage in real-time.
Traditional LLMs generate tokens one at a time. Diffusion LLMs generate tokens in parallel — refining the entire response at once. At 150ms response times and 2000+ tokens/second, UI interaction feels instant. And it's getting faster every quarter.
This extension explores what becomes possible when language models are fast enough for real-time UI manipulation.
- Clone the repo:
git clone https://github.com/your-username/dllm-lens.git - Open
chrome://extensionsand enable Developer mode - Click "Load unpacked" and select the
extension/folder - Click the extension icon and paste your API key
| Model | Provider | Speed | Default |
|---|---|---|---|
| Mercury Coder | Inception Labs | ~2000 tok/s | Yes |
| Celeris-1 | Celeris AI | ~2000 tok/s | |
| Any OpenAI-compatible | Various | Varies |
BYOK — bring your own key. The extension calls the model API directly from your browser. No backend server, no data collection.
Want to add a model? See Adding a Model Profile.
Guide mode — Ask where something is on the page. The extension analyzes the DOM, finds the element, and highlights it with an overlay.
"Where do I change my shipping address?"
Simplify mode — Tell it something is confusing. It rewrites the DOM region to be clearer, preserving functionality. Fully reversible.
"Simplify this checkout form"
Under the hood: the extension prunes the page DOM (strips scripts, styles, hidden elements, collapses wrapper divs), sends it to a diffusion LLM with your request, and applies the structured response — either as a visual overlay or a DOM rewrite.
- Large or deeply nested pages may hit model context limits
- Rewrites are cosmetic — they don't persist across page loads
- Complex SPAs with heavy client-side rendering may produce stale DOM snapshots
- No vision input yet (DOM text only) — accuracy depends on DOM quality
- Some sites with strict Content Security Policies may block the widget injection
- Vision input — send screenshots alongside DOM for better element identification
- Streaming rewrites — progressive DOM updates as the model generates
- Automatic improvements — learn from user interactions to suggest fixes proactively
- Site-owner embed mode — drop a script tag to improve your site's UX for all visitors
The most useful contributions right now:
- Try it on unusual sites and file issues when it breaks
- Add model profiles for new diffusion LLMs as they launch
- Improve DOM pruning — better heuristics for what to keep and what to strip
MIT