Semantic Cmd-F

Find anything on a page by meaning, not just keywords.

Ctrl+F has worked the same way for 30 years: it finds the exact letters you typed. Semantic Cmd-F is the upgrade. Press Ctrl+Shift+F on any page, type what you're looking for in plain English — "the part where they explain the refund policy", "pricing details", "where they criticize remote work" — and it scrolls you to the most relevant section. No exact keywords required. It works because it runs a small open-source AI model entirely inside your browser. Your queries and the contents of the pages you read never leave your computer. There's no server, no account, no analytics. The whole thing is local-first by design. WHEN IT'S USEFUL • Researching: jump to the right paragraph in a 10,000-word article • Studying: find the part of a textbook chapter that explains a concept, even if the textbook uses different words than you do • Reviewing: locate a specific clause in a long terms-of-service or contract • Reading code documentation: find the section that answers your question, not the one with the matching keyword • Searching long blog posts, papers, or transcripts where you remember the idea but not the wording HOW IT WORKS 1. Press Ctrl+Shift+F (or Cmd+Shift+F on macOS) on any web page 2. Type a description of what you're looking for in natural language 3. The extension scrolls to the most relevant section and highlights it 4. ↑/↓ to navigate between the top matches; Esc to dismiss PRIVACY-FIRST • No data collection. Period. • No accounts. No sign-up. • No analytics, no telemetry, no tracking pixels. • No advertising. • The first time you use it, the extension downloads about 22MB of model files from Hugging Face's public CDN, then caches them locally. After that, no network requests are made. LIMITATIONS (BE HONEST) • First search takes a few seconds while the model downloads. Subsequent searches are fast. • Works best on pages with normal article markup (paragraphs, headings, lists). Pages built entirely from generic <div> tags may produce uneven results. • Searches the currently-visible page only — no cross-tab or history search in this version (planned for v2). • Currently English-tuned; the underlying multilingual model handles other languages but quality may vary. Source code and detailed privacy policy linked below.