Page-aware AI

How Cordy grounds answers in the page you're reading — reading page content on demand, what's sent, the confirm dialog, and the safety model.

Page-aware AI is what lets Cordy answer questions about the page in front of you. Instead of copy-pasting an article into a chat box, you ask about "this page" and Cordy reads the readable content of the current tab and grounds its reply in it — on demand, and only when you ask.

How it works

Cordy never watches your browsing in the background. Page content enters a conversation only when you pull it in, in one of these ways:

  • Ask with the page attached — @-mention the current page (see Mentions) or use a page action, and Cordy reads the current tab's text as context for that message.
  • The model requests it — during a chat, the assistant can call the getTabContent tool to read the current page. This is a public-read tool and runs automatically when browser tools are enabled; tools that expose broader browser data or change browser state use the approval policy described in the Tools guide.
  • From the floating menu — act on a selection instead of the whole page (see Floating menu).

Either way, reading a page requires you to have granted site access. Without it, Cordy tells you to enable In-Page Features in Settings → General rather than silently failing.

What Cordy reads

When a page is used as explicit mention context, Cordy extracts its readable text — the main content and title, rather than raw HTML, scripts, or hidden markup. Extraction stops at 12,000 characters or 8,000 visited text nodes. The outgoing mention block has a separate model-window-aware budget, including a 5,000-character ceiling for each reference, so the text delivered to the model can be shorter. The confirmation dialog reports the delivered amount.

For multi-tab requests (for example, "compare these tabs"), Cordy reads the readable content of the tabs you mention, up to a sensible limit, and labels each source.

The mention confirm dialog

Before sending context that you attached with @ to a cloud model, Cordy shows a "Send browser context?" dialog. It spells out:

  • what's included — the links, titles, and readable page content you attached,
  • roughly how many characters of content are included, and
  • whether a reference is full content or metadata only (title and URL).

Approve it to send, or cancel to keep the context local. After a reply, Cordy notes how many context sources were actually used.

This is separate from the one-time cloud consent dialog, which you accept only once. The mention confirm dialog is per-send and covers the context you attached. A getTabContent tool call follows its public-read tool policy and does not show this mention dialog; its result becomes context for the provider request that continues the tool turn.

What you can do with it

  • Research — summarize the article you're reading, pull out key claims, or compare information across several tabs.
  • Development — explain API docs, decode an error message, or walk through code on the page.
  • Shopping — compare specs, digest reviews, and weigh alternatives from the pages you have open.
  • Learning — get a plain explanation of a dense page, or turn it into questions to test yourself.

Many of these have ready-made prompts under the chat's browser-context workflows: Summarize selected tabs, Compare tabs, Cleanup plan, and Save conclusion — each acts only on the tabs you explicitly mention with @.

Safety: page content is untrusted

A page you're reading can contain text that tries to hijack the assistant ("ignore your instructions and…"). Cordy defends against this prompt-injection risk by treating all browser-derived data — page body, selection, and tool results — as untrusted input, kept separate from your actual instructions. Page content informs the answer; it doesn't get to issue commands or silently trigger destructive tools (those always need your approval — see Tools).

Privacy summary

  • Page content is read only when you ask — never collected in the background.
  • Reading pages requires site access, which you grant and can revoke.
  • Context attached with @ goes to the configured cloud provider only after the mention confirmation. Browser-tool results follow the tool policy described above. Local, on-device models keep their processing on your machine.

Full details are in Privacy & permissions.