browsa

browsa is an AI sidebar that chats with the page you're on — articles, videos, PDFs, and docs — using YOUR own key or YOUR own agent. No subscription. No account. No middleman. ────────────────────────────── 💬 CHAT WITH ANYTHING YOU'RE READING • Attach the current page with one click — clean reader-mode extraction, not a DOM dump. Interactive expansion runs on a hidden background copy, so it never clicks around the page you're reading • Summarize, translate, explain, outline — one-tap quick actions and slash commands • Select any text on the page → floating toolbar to ask / explain / translate / summarize 🎬 VIDEO NOTES WITH CLICKABLE TIMESTAMPS • Subtitles of the video you're watching are extracted automatically • No subtitles? browsa transcribes the audio (ASR) through your own configured provider, so subtitle-less videos work too • Every [mm:ss] note in the reply is clickable and seeks the video 📄 PDFs & OFFICE DOCS NEVER LEAVE YOUR MACHINE • Full layout reconstruction client-side (tables, headings, multi-column) — powered by WASM, zero upload • Office documents convert to clean Markdown the same way — headings, lists and tables keep their structure • Charts and figures are cropped and sent as images so vision models can actually see them 🤖 BRING YOUR OWN MODEL — OR YOUR OWN AGENT • Any OpenAI-compatible endpoint: hosted APIs, local gateways, self-hosted deployments — with a per-model thinking level (deeper reasoning or faster replies) • Local & offline models via Ollama and other local runtimes • Agent backends too: self-hosted servers (e.g. Hermes) with tool execution, streaming progress, and human approval before risky actions • OpenSquilla, a local desktop agent, connects through its gateway — sessions and memory stay on your machine • Or drive coding agents already installed on your machine — CLI or desktop app — through a local bridge: each agent signs in with its own configured account, no separate model API purchase • Connect several agents at once: the sidebar dropdown switches between them, every reply is labeled with the agent that produced it, and each agent keeps its own connectivity status • Cross-entry continuity: agent sessions are named after your conversation automatically, and the sessions drawer shows a copyable session ID — pick the same thread back up in your agent's own CLI • Multiple wire protocols supported, so almost any provider or gateway plugs in 🌐 DEEP SITE EXTRACTION (not just "grab the text") • Popular video, community, podcast and document sites are extracted through their own structured data — cleaner and more complete than generic page scraping • GitHub file and repository pages fetch the raw source instead of the rendered UI 🔒 PRIVACY BY ARCHITECTURE • Your API keys are stored only in your browser's local storage and sent nowhere except the endpoint you configure • Page context is sanitized locally before sending: credentials hiding in URLs (tokens, passwords, signatures) are masked so they never reach your provider • PDF parsing runs entirely on-device (WebAssembly) — file bytes never leave your machine • Model replies are sanitized (DOMPurify) before rendering; network interceptors only observe requests, never modify them • Open source (MIT): every line is auditable — https://github.com/xiaohuzai/browsa ✨ RICH REPLIES, RENDERED PROPERLY Markdown tables, code highlighting (40+ languages), math formulas, flowcharts, data charts and mind maps — rendered inline. So are molecules and chemical reactions — molecule structures are validated before drawing (impossible chemistry is flagged, not silently drawn) with a properties line (weight, logP, TPSA…) — and interactive 3D protein structures with confidence coloring, fetched by ID from public structure databases and drawn with the same viewer engine those sites use. Neural-network & graph figures are laid out by Graphviz; diagrams, charts and 3D viewers are drag-resizable and adapt to the dark theme. Code you type into the input box is highlighted the same way. Formulas are copyable: select any reply to copy clean LaTeX source, or hover a formula for a one-click copy button. Select any part of a reply to open a scoped side-conversation ("detail thread") — you can paste images into it too. ⌨️ MADE FOR KEYBOARD PEOPLE Ctrl+Shift+H opens the sidebar · Enter/Shift+Enter configurable · Ctrl+F search inside the conversation · Ctrl+/ switch context mode · everything streamable and abortable with Esc ────────────────────────────── browsa does not add markup to anyone's model bill — you pay your provider directly, or run local models for free. It collects no analytics and requires no sign-up. The UI follows your browser language — English, 中文, 日本語, 한국어, Español, Português, Русский — switchable in Settings. Install, open ⚙ Settings, paste a key (or point at localhost:11434), Ping to verify — you're chatting in under a minute. Source code & docs: https://github.com/xiaohuzai/browsa Website: https://xiaohuzai.github.io/browsa/