Recommended · Automatic updates
Install from the store
The simplest way to start. Add the extension, then connect your model in Settings. Store review may lag behind GitHub releases.
Add to Chrome ↗browsa / browser side-panel AI
Stay on the page.Ask beside it.
Bring the page into a conversation with your own AI — no copy-paste, no leaving what you’re reading.
Chrome / Edge · Source install available
Free extension. Bring your own model or agent.
▶YouTube
Search…
4K
Blue Planet II · Episode 1 – One Ocean (full documentary)
From the shallow seas to the abyss, explore the most hidden corners of the ocean:
bioluminescent creatures, cooperative humpback hunting, and life in total darkness.
We know less about our ocean than the surface of the Moon.
18:42Blue Planet II · Ep.2 – The Deep
The Deep · 4.1M views
21:07Ocean Bioluminescence Explained
Bioluminescence · 890K views
15:30How Humpbacks Hunt with Bubble Nets
Bubble-net hunting · 2.3M views
📎 Attached: 《Blue Planet II》Ep.1 – One Ocean
/summarize key points, with timestamps
Reading subtitles…
One Ocean · key moments
🌊 Intro on the ocean's scale — we know it less than the Moon[00:12]
✨ Bioluminescence is widespread in the deep, for lures & camouflage[03:40]
🐋 Humpbacks hunt cooperatively with bubble nets[08:15]
🐙 Abyssal life evolved extreme adaptations to pressure and dark[11:48]
🌍 Close: human activity is reshaping ocean ecosystems[15:02]
Ask anything… use 📎 to attach current page
Video notes: replies carry clickable [mm:ss] timestamps — click to jump to that moment.
AI · Browsers
Bringing an AI copilot to the side of every page you read
Feeding a page's text to an LLM is easy; making it truly "understand" the page is hard. We use a three-layer cascade: Readability first for clean article text, then a DOM tree that keeps structure, then full text as a fallback. For major sites we read the structured data the page already loaded — subtitles, comments, article source.
No re-authentication, no third-party uploads — data flows only between your browser and your own backend.
📎 Attached: Bringing an AI copilot to every page you read
/keypoints what are the takeaways?
Reading article…
Takeaways
🧠 Three-layer cascade: Readability → DOM tree → full text
🕵️ Major sites read their own loaded data — no re-auth
🔒 Data stays between your browser and your backend
⚡ Articles, subtitles & PDFs all covered
Ask anything… use 📎 to attach current page
Articles: /keypoints extracts the essentials; the body is attached via the Readability cascade.
Attention Is All You Need
Abstract — The dominant sequence transduction models are based on complex recurrent or convolutional neural networks… We propose a new simple network architecture, the Transformer, based solely on attention mechanisms.
The Transformer dispenses with recurrence and allows for significantly more parallelization, cutting training time.
Figure 1: The Transformer – model architecture
It achieves state of the art on WMT 2014 English-to-German and English-to-French at a fraction of the training cost.
📎 Attached: PDF 《Attention Is All You Need》(15 pages · 6 figures)
What's the core contribution?
Parsing PDF…
Core contribution
⚙️ Transformer — attention-only, no recurrence or convolution
⚡ Highly parallel, trains far faster than RNN/LSTM
📊 Fig. 1 shows the architecture; Fig. 2–4 training & BLEU resultsFig.1
🏆 SOTA on both WMT 2014 tasks at lower training cost
Ask anything… use 📎 to attach current page
PDFs: local WASM rebuilds tables / headings / layout; figure regions are cropped and captioned automatically.
Illustrative interface · Switch scenes to see how browsa works with the page.
demo
Stay on the page. Ask beside it.
The full workflow in 68 seconds: attach a document → math & diagram rendering → clickable timestamps → session management and Agent handoff.
how it works
Three steps to talk to a webpage
No window-switching, no copy-pasting. browsa sits quietly beside the page you're reading — like a margin note.
01
Open the side panel Ctrl+Shift+H
Press the shortcut or click the toolbar icon on any page. The browsa panel appears on the right.
02
Attach the page 📎
Click 📎 and the page content flows into the conversation — articles, subtitles, PDFs. Selected text can be asked about directly.
03
Stream replies any backend
Replies stream in word by word, with Markdown, math and charts. Any OpenAI-compatible API — or the Hermes / opencode / OpenSquilla / agent-bridge agents. Seven UI languages included.
Installed? Quick start walks you through the first question; the full guide covers reading, video, rendering, sessions, backends and settings.
features
More than a chat box
browsa puts "reading" and "asking" in the same place: long articles, videos, PDFs — the AI is right there.
attach
Attach the page
Bring the current page into the conversation with one click — modes pick themselves:
- Auto cascade: Readability body → DOM tree → full text
- 📷 Screenshot mode, for multimodal models
- PDF parsed locally: tables / headings / layout rebuilt
- GitHub file pages pull raw source directly
select
Selection Q&A
Select any text on any page and ask directly:
- Floating toolbar: Ask · Explain · Translate · Summarize
- Explain / Translate stream inline, without interrupting your reading
- Ask / Summarize ride into the panel — no 📎 needed
render
Rich rendering
Replies are more than plain text — formulas and diagrams render in your browser:
- Markdown GFM + 40+ language syntax highlighting
- LaTeX math (KaTeX, Web Worker accelerated)
- Mermaid flowcharts (one-click horizontal/vertical switch, re-layout on panel resize, keyboard/two-finger zoom), ECharts, Markmap mind maps
- Broken Mermaid? One-click AI fix and re-render
video
Video notes
Structured subtitles from YouTube and Bilibili:
- Clickable
[mm:ss]timestamps that seek the video - Subtitle drawer: live highlight · click to seek · note a line
- No subtitles? ASR transcription / A/V deep-read
- Xiaoyuzhou podcasts: one-click transcription with timestamps
- Long transcripts auto-compressed into points, timestamps preserved
provider
Any backend
No vendor lock-in — connect your own models:
- Hermes Agent: tool execution, approval for risky actions, progress display
- OpenCode Agent: direct CLI line — full tool & approval flow
- OpenSquilla Agent: gateway WebSocket — server-side session memory
- Agent Bridge: bridges Codex / Claude Code / Gemini CLI / pi / dsh / ZCode — one card, several agents
- Agent sessions live agent-side — auto-named, easy to resume
- Any OpenAI-compatible API: OpenAI, Ollama, LiteLLM…
- Multiple models per card — switch right in the sidebar
- Per-model thinking level, request fields auto-adapt
- Ping to verify connectivity and capabilities
privacy
Privacy first
No browsa account. Your choice of model service:
- API keys are stored locally and used to authenticate with your configured provider
- Locally masked before sending: URL tokens / passwords / signatures
- PDFs parsed on-device — content never leaves the machine
- Office documents convert to Markdown on-device
- Content goes only to the backend you configure
input
Fast input
Say more without typing more:
- Quick actions: Summarize / Key Points / Explain / Outline
- Slash commands + ↑/↓ input history; drafts survive closing the panel
- Keep typing while a reply streams — messages queue and send in order
- Pasted code becomes a highlighted block; everything else stays byte-for-byte
session
Sessions
Long conversations stay organized:
- Save / restore / export Markdown, with Agent thread IDs
- Pin sessions — searchable by title and full text
- Switch sessions mid-reply — it finishes in the background
- Full-text in-conversation search (
Ctrl+F) - "Detail-thread" any text without polluting the main chat; paste images too; the selection stays highlighted and the card folds; reopening the panel restores it folded without losing anything
supported sites
It reads your webpages
For these sites browsa reads the structured data the page already loaded — subtitles, comments, article source, quotes — instead of guessing.
YouTubesubtitles/chapters
Bilibilisubtitles/chapters/ASR/stats
Xiaohongshunotes/photos/comments
Juejinarticle source/images
Zhihuarticles/answers/images
X / Twitterposts/photos/replies
GitHubrepo metadata/README
Redditposts/comments/photos
Xueqiuquotes/posts/images
Xiaoyuzhoupodcasts/ASR
Dedaoarticles
GeekTimearticles
Observes, never hijacks: it just watches the requests the browser itself makes. Every other webpage works too, just not as deeply.
your keys, your agent
Your keys. Your models. Your agent.
browsa sells no subscription, asks for no account, takes no markup. It simply moves your own model endpoints into the side panel — cloud, local, or self-hosted.
- ›No account, no telemetry. Install and go — nothing ever flows through us.
- ›API keys stay local, sent only to the endpoints you configure.
- ›Ollama / LM Studio local models — fully offline works too.
- ›Real agents: tool execution, live progress, approval before risky actions — with an optional chime when one arrives (toggle in Settings).
O
My OpenAIgpt-4o · /v1/chat/completions
reachable
C
Claudeclaude-sonnet · /v1/messages
not pinged
L
Ollama (local)localhost:11434 · qwen3:32b
offline ok
H
Hermes Agent/v1/runs · tools · approvals
reachable
oc
OpenCode Agentopencode serve · tools · approvals
not pinged
Sq
OpenSquilla Agentgateway /ws · tools · approvals
not pinged
↔
Agent Bridgecodex / claude / pi / gemini / dsh / zcode · connect local agents
bridge
Not sure which to choose? Read the model and agent setup guide →
install
Start with your next read.
Choose one installation method, then connect your model or agent. The browsa extension is free.
Installed? Make the connection.
Configure a model API or local agent in Settings and Ping to verify. Open an article, attach it with 📎, and ask your first question.