Reading pages

"Reading" is browsa's foundation: one click on 📎 brings the page you're on into the conversation. This page covers what it reads, how, and the other ways in besides 📎.

📎 Attach the current page

Click 📎 next to the composer and the current tab is attached to the conversation. The default Auto mode picks its reading strategy in a cascade: clean article text first (Readability), falling back to the structured DOM tree, then to full page text. You never choose — the note above the composer tells you which reading was used.

One attach, lasting context: every following turn carries the page content with it, until you clear history or re-attach a different page.

📷 Screenshot mode

Sends the visible area of the tab as an image to a multimodal model — for questions plain text can't answer, like "which curve rises here". Toggle with Ctrl+/ or pick it in 📎's mode menu.

Select text: inline explain / translate

Select text on a page and a floating toolbar appears on release: Ask · Explain · Translate · Summarize.

• Explain and Translate answer in place: a streaming card opens right next to the selection — pronunciation and examples for a single word, no side panel needed, read and move on;
• Ask and Summarize ride into the side panel automatically — they start a conversation;
• right-click → browsa › Ask / Explain / Translate / Summarize sends all four to the panel.

Toolbar and right-click labels follow the UI language in Settings. Selection answers use an explicit reply language first; in Auto, explanations and summaries use the UI language and translations target it. When the selection is reliably identified as already in that language, Auto switches the translation target; an explicit reply language keeps its selected target.

None of these need 📎 first.

Images

Drag an image into the composer or paste a screenshot with Ctrl+V (multimodal models only — GPT-4o, Claude, and other vision models).

PDF: parsed on your machine, figures included

Attach a PDF — or open a page that turns out to be one — and no mode picking is needed:

• Parsing happens entirely in your browser (WASM + pdf.js); the file's bytes never leave your device;
• Tables, headings, and multi-column layout are reconstructed — the model reads structure, not scrambled text;
• Figure regions are cropped out and sent as images to vision models, paired with their captions — the model can actually see Figure 1, and the reply cites it as [图N] (same contract as web article images).

GitHub repos & file pages

On github.com file pages, browsa fetches the raw source from raw.githubusercontent.com directly — Markdown and code keep their structure instead of the rendered page. On a repository page, the attach is structured metadata (description, topics, license, stars) plus the raw README — the About sidebar stops being discarded as nav chrome.

Deep extraction on known sites

For the sites below, browsa observes the requests the browser itself makes and reads the page's already-loaded structured data — subtitles, comments, note text — rather than guessing; when interception misses, the site's own API is queried on demand. No re-authentication, nothing sent to third parties.

Pages not on the list work too — via the Auto cascade, just not as deep.