Browser

The DeutschlandGPT extension puts a chat in your browser's side panel. When you switch agent mode on, it also lets that chat act on the page you are looking at: read it, click, type, and move between tabs. This page explains what it may do, and where you control it.

Which browsers

Supported: Google Chrome 116 and later.

Microsoft Edge and other Chromium browsers can install the extension from the Chrome Web Store and may work, but are not officially supported.

Safari and Firefox are not supported, because they do not run Chrome extensions.

Why you would want it

A chat without the extension only knows what you paste into it. With the extension it knows the page in front of you. That is the signed-in page, so what sits behind a login too, which a link alone could never show it.

That changes three kinds of task in particular:

• Reading instead of copying. A long ticket, a contract in a customer's portal, a table in an internal tool: "summarise this", without copying anything out first, and without the page needing to be publicly reachable.
• Doing instead of describing. With Agent mode on it can click, fill fields, follow links and switch tabs. "Here is how you fill in that form" becomes "fill in that form". Every step that changes something is asked first.
• Across several tabs. It can open and switch tabs, so it can look something up and put the answer where you are working.

What the chat additionally sees

Only what you hand it, and only while the conditions below hold:

• The content of the tabs in its tab group: the tab you opened the panel on, tabs it opened, and tabs it took over: the page's text and structure.
• Screenshots of those tabs, when it needs to look at something the text does not carry.
• The titles and addresses of all your open tabs, so it can tell which one you mean when you say "the other tab". It reads the content of a tab only once that tab is in its group.

Not: your history, your bookmarks, your saved passwords.

Four things have to be true before it touches a page

The extension is the most far-reaching permission the product asks for: the browser you are signed into, with every session you have open. That is why it is deliberately behind more than one switch:

1. The extension is installed and you signed in through it. That is what pairs the browser.
2. Agent mode is on for that chat. Without it the panel is an ordinary chat; it can talk about the page only once you allow it to read one.
3. Your organisation has the extension enabled. If it does not, the settings page does not exist for you.
4. You have not switched the connector off on Settings → Browser.

Any one of them being false is enough to stop it.

Which DeutschlandGPT surfaces can drive a paired browser: the web app, the desktop app and the iPhone app. The Android app and the Office add-ins cannot. They have no way to show you the question a changing step asks, so they are not offered the browser at all rather than stopping halfway through one.

What it asks before doing

Anything that changes the page (a click, typing into a field, following a link) asks you first, in the panel, with the action named. Reading the page and taking a screenshot do not ask each time, because a chat about the page you are on would otherwise be a wall of confirmations.

You can loosen or tighten that per chat with the approval control in the composer. On an approval card, Allow for this chat ends with the chat; Always allow applies to every chat and is listed under Permissions → Browser, where you can take it back. This is the one control there is: no site and no kind of input asks on its own account. If you choose to skip approvals, the extension acts without asking everywhere, including on banking and payment sites and on sign-in pages.

Where it cannot work

The extension blocks no website. Two kinds of page remain out of reach:

• DeutschlandGPT itself: the extension does not act on the app you are chatting in.
• Your browser's own pages (chrome://, edge://, extension pages, the Chrome Web Store): Chrome lets no extension read or act there.

On either one, the tab can still be in the group; the assistant says it cannot work there, and carries on as soon as the tab is on an ordinary page again.

Which tab it can act on

Only tabs in the DeutschlandGPT tab group. Opening the panel always puts the tab you opened it on into that group, on any page. Beyond that tab, the group holds the tabs the assistant opened itself and any of your open tabs it took over because the task needed them. It takes a tab over without asking; the chat shows "Took over tab" with the tab's title, and the tab appears in the group. You see the group in Chrome's tab strip at any time.

Opening a tab yourself does not put it in scope, even from a tab that is. Dragging a tab out of the group takes it away from the assistant. It stays out until you share it again: with the toolbar icon, or with "Share this tab" above the input field. Reloading the panel does not bring it back.

Taking it back

Settings → Browser lists every browser you have signed into, with when it was last seen:

• Revoke access disconnects that browser. The extension stays installed; signing in again reconnects it.
• The switch at the top stops chats reaching any of your browsers, without uninstalling anything. The side panel keeps working as a chat.

Revoking is per browser and takes effect on the next turn.