Skip to content

The browser agent

The browser agent is Fadenstack in Chrome's side panel. It works with the page you have open: it answers questions about it, summarises and translates it, shows you where something is, copies tables out, compares tabs, reads PDFs and video transcripts, looks at screenshots, writes into the mail or ticket you are working on, and goes through a form over several pages. It answers with your organisation's model, through your Fadenstack server, and in the Ask tool mode, where every chat starts, it asks before it changes anything.

The browser agent's panel: suggested questions, the bar naming the open page, and the message box with its buttons below

Before you start

  • The extension is installed in Chrome. Your IT department installs it for you, or tells you where to get it.
  • Access: your administrator has given you the browser agent. See Agents and add-ins.
  • The server's address and the agent's name, unless your IT department set them for you.

Open the panel and sign in

  1. Select the Fadenstack button in Chrome's toolbar, or press Alt+Shift+F.
  2. The first time, choose Connect to my organisation (your IT department may have chosen for you), then enter the Server and the Agent in Connect to Fadenstack.
  3. Select Connect, and allow the access when Chrome asks.
  4. The panel shows Your sign-in code. Select Open the sign-in page; your browser opens Fadenstack's page for it. If you cannot open it from the panel, Copy the code copies the code.
  5. On that page, sign in to Fadenstack as usual if the browser asks, check that the code is the one in the panel, and approve. The panel says Waiting for your approval... meanwhile and signs in by itself afterwards. Your password stays on that page; the extension never sees it. More: Sign in a desktop app.

If your IT department set the server and agent, the form shows them and you only select Connect and sign in.

Sign in with a password instead

Select Sign in with a password, enter your E-mail and Password (your Fadenstack account), and select Sign in. This is also the way in when your server is older and cannot sign in with a code: the panel says This server cannot sign in with a code yet. Sign in with a password. Sign in with a code brings the code back. If the code expired or was denied, select Get a new code.

Server and Agent
  • Server: your Fadenstack server, for example https://ai.example.internal.
  • Agent: its address, as the console shows it under Agents; browser-assistant unless you were told otherwise.

Language and settings

The panel shows itself in your browser's language. To pick another, select the gear (Settings) at the top of the panel, then a language under Language: the whole panel changes at once, the conversation, the agent's greeting and the right-click items too. Automatic goes back to the browser's language. The language is kept for your account, so the console and the Office add-ins follow it, on this and your other computers. If your organisation set a language, the settings show it with Set by your organisation, and you cannot change it.

Below the languages, Settings holds the switches Search the web with Chrome and Use tools that sites offer (experimental), and shows which agent you use on which server, with these items:

Select What happens
Your data in this browser Opens the page for your data. Export my data saves a JSON file with the chats, saved prompts, Allow always choices and record of tool calls kept for you in this browser. Your sign-in is not in it, and neither is what the server keeps; ask your IT department for that. Remove my data… asks you to confirm, then removes the chats, saved prompts, Allow always choices and record of tool calls kept for you in this browser, and ends your connection, so you connect again from the start. This cannot be undone. Your chats cannot be read again in this browser. Export all my Fadenstack data and Remove all my Fadenstack data in this browser… do the same for every Fadenstack app in this browser. What the server keeps stays; ask your IT department.
Connection Shows the server you are connected to. Sign out ends your sign-in at the server and on this computer; tick Also delete the chats kept on this computer to delete your chats too, otherwise they stay in this browser for your next sign-in. Here you can also change to another server or to your own AI.

The chats and choices are kept for each signed-in person on their own: someone else who signs in to the same browser does not see yours.

Settings also holds Tools, which lists what the agent can use (see Tools), and Export chat, which saves the open chat as a Markdown file.

About, the last entry under the gear, names the agent and its version, the model, the extension, the Fadenstack SDK it is built with and the server; Copy details copies them for a question to your IT department. It links to your organisation's Help and to fadenstack.com. A small Fadenstack link also sits under the message box (what such a link sends: Your data and privacy).

Ask about a page

Type your question in the panel, for example "Summarise this page" or "What does this contract say about notice periods?". The bar above the message box names the page the agent works with.

What the agent reads on a page

The agent reads the page itself when you ask about it: its text, headings, lists and tables, the visible part, without menus, footers and side bars, and without text the page marks as hidden. A long page is read in parts. Parts of the page that come from its own site in a frame, such as an editor or an embedded form, are read in place; a frame from another site is read when you have allowed that site too.

The buttons under the message box add something to your next message:

Button Adds
Add this page (the page) The page you are on
Add the selection (the quote mark) What you selected on the page
Add tabs (the tab) Tabs of this window you pick, to compare them (All readable tabs takes all of them)
Add a PDF from this computer (the paper clip) A PDF file; you can also drop it on the panel. With a model that sees pictures: Add a PDF or a picture from this computer
Add a screenshot (the camera) What the page shows, as a picture. Only with a model that sees pictures
Quick actions (the sparkle) See Quick actions

Add tabs opened: All readable tabs and the window's two tabs to tick, with Cancel and Add

New chat at the top of the panel starts a fresh conversation; Chat history lists your earlier ones in this browser.

More about screenshots and pictures

Pictures go to the agent only when its model sees pictures; otherwise the camera is not shown, and a picture you paste is not added: the panel says so. You can also paste a picture into the message box, or drop one on the panel.

Add a screenshot takes the page directly when you opened the panel from the toolbar button, the shortcut or a right-click item on that page. Otherwise Chrome asks what to share: pick the tab, another window or your screen. Fadenstack takes one picture of it, and the sharing ends at once.

Quick actions

Quick actions ask the agent at once, in your language. They are in the panel's Quick actions menu and in the right-click menu under Fadenstack:

On Quick action
The page Summarise this page, Translate this page into …, Ask about this page, Ask about a screenshot of this page
Text you selected Explain, Summarise, Translate into …, Ask about "…"
Text you selected in a field or editor Improve the writing: the agent's version replaces your selection, after you allow it

Translate into … offers the panel's language and English. To ask about a selection from the keyboard, select the text and press Alt+Shift+S: the panel opens with the selection added.

Skills and your own prompts

Skills are ways of working your organisation set up, such as "Reply in our style" or "Check a contract against our checklist". To use one for a message, type $ in the message box (or / at its start), and pick the skill from the list: type a few letters to narrow it, then press Enter. The skill shows above the message box and goes with that message only. See Skills.

The same list starts with My prompts: questions you saved yourself. To save one, select Save as prompt (the bookmark) under a message you sent. Picking a prompt puts its text into the message box, where you can change it before you send it. The trash icon beside a prompt deletes it. Your prompts are kept in this browser only, up to 100; Export my data includes them, and Remove my data… deletes them.

Videos

On a page with a video, the agent can read its transcript: ask "Summarise this video" or "What does it say about the warranty?". It reads the captions the video has, on YouTube also the transcript YouTube shows, each line with the time it is said at, and quotes the times. Ask "Where does it say that?" and it moves the video to that point; you can also give a time, such as "Go to 3:42".

When a video has no transcript

A video without captions or a transcript cannot be read; the agent says so. Captions made by speech recognition may mishear names and numbers: check what matters in the video itself.

Tasks across pages

The agent can go through a task that spans several pages of a site, such as a form in steps or a search on the site: it goes to another page of the same site, goes back, scrolls, waits for the next page, and presses Enter, Tab, Esc or the arrow keys in a field. While it acts on a page, the page shows a coloured edge and, at its foot, Fadenstack is working on this page with a Stop button: Stop there stops the answer as the panel's Stop does.

It stays on the site the tab is on. Another site it opens in a new tab with Open page, which needs your permission for that site, or reaches by clicking a link, which asks you. One request takes at most 25 actions on pages; then the agent tells you what is done and asks whether to go on. It is told never to buy, pay, create accounts or accept terms for you, and to stop before such a step and hand it to you.

Tables and chats as files

Every table in an answer has Download CSV beside its copy button: the table as a file for Excel or another spreadsheet, saved where Chrome saves downloads. Export chat, under the gear, saves the open chat as a Markdown file, with your messages, the names of what you added (not their content) and the agent's answers.

Cells that start with = + - @

In the CSV file such a cell starts with an apostrophe, so a spreadsheet shows it as text instead of running it as a formula.

Tools

Tools, under the gear, lists what the agent can use. In this app: the extension's own tools, each with what it may do (Read, Changes data, Cannot be undone). From your organisation's server: the tools of your organisation's connected services, grouped by service, each marked available to this agent, switched off by your administrator, or not allowed for this agent.

Tools that sites offer (experimental)

Some sites offer assistants actions of their own (WebMCP), such as "find an order" on a shop's page. Chrome tests this standard and offers it only in some versions or on some sites. The agent uses such tools only when you switch them on: select the gear, then Use tools that sites offer (experimental). If your organisation decided for you, the item shows Set by your organisation.

Every use of a site's tool asks you first, with Allow and Deny only, whatever the site says about the tool: the site runs it, and it may change something there. What a site says about its tools, and what they answer, the agent treats as the site's text, never as instructions.

The agent can search the web with Chrome's own search engine, the one set in Chrome. It is off until you switch it on: select the gear (Settings), then Search the web with Chrome. If your organisation decided for you, the item shows Set by your organisation.

In the Ask tool mode, every search asks first, and the card shows the query. The search runs in a tab in the background, and the agent reads the results like a page, so the first time Fadenstack needs your permission for the search engine's site: select Allow … above the message box, then ask again. It then opens a result or two and gives you their links.

Where a search goes

The query goes to the search engine set in Chrome, not through your organisation's server, as a search you typed into the address bar would. Your organisation's server may still see the query: where it supports this, it checks each step, with what the step is given, before the step runs. The agent is told never to put personal data or text from the page into a query unless you asked for exactly that; read the query on the card before you allow it.

What else it can do

Ask it to What it does
Show you where something is on the page It marks the passage and scrolls to it. In a PDF it opens the PDF at the right page; in a video it moves the video there.
Copy a table or text It puts it on your clipboard as text and as a table, so it pastes into Excel as cells and into Word or Outlook as a table. Every table in an answer also has a copy button.
Write into a mail, a ticket or a form It fills form fields, and writes formatted text into the editor on the page, such as a mail in Outlook on the web or Gmail, a ticket, or a note. Say where: at your cursor, below the greeting, at the end, or replacing everything.
Improve or translate text you selected in a field It puts its version in place of your selection.
Answer from your organisation's documents It searches the knowledge your administrator gave the agent, and names the document it used. See Knowledge.
Read PDFs It reads PDFs open in Chrome, and PDFs you add. Long PDFs are read page by page as needed.
Look at a picture With a model that sees pictures: screenshots and pictures you add, such as a chart or an error message.

What it asks first

The Tool mode at the top of the panel decides what runs without asking:

What Off Read only Ask (default) Auto
Read the page, tabs, PDFs, a video's transcript, the selection; search; show where; scroll; wait for a page no runs runs runs
Fill a field, write into an editor, replace your selection, copy to the clipboard, open a page, search the web, go to another page of the site, go back no no asks runs
Press a button or follow a link, press a key (Enter may send something), use a tool a site offers no no asks, every time asks, every time

A card asking Run copy_to_clipboard?, showing the table it will copy, with the Allow, Allow for this chat and Deny buttons

When it asks, a card shows Run …? with the action and what it is given, and says what it acts on, such as Button "Send" on mail.example.com, Press Enter in Text box "Search" on shop.example or Go to https://shop.example/order/2. Select Allow (this once), Allow for this chat (until you close or reopen the chat), Allow always (for this agent in this browser), or Deny. Pressing a button or a key, using a site's tool, and anything your administrator wants asked first, offers only Allow and Deny: it asks every time, except in Bypass. Reset approvals, at the bottom of the tool-mode menu, takes back what you allowed for the chat and always.

Where your organisation allows it, Bypass is a fifth tool mode. You switch it on yourself, for one chat and a limited time. Then what your organisation's Bypass grant covers runs without asking, a button or a key press included; rules your administrator set, and tools marked "always ask", still ask. See Bypass.

What the card shows for a field or a button

For a field or a button, what the action is given is the page's internal reference, not its label.

If you close the panel while the agent answers, the answer stops: the agent works in the panel. When you open the panel again in that window, the chat is back, says The answer stopped when the panel was closed, and Retry asks again.

What it writes is marked

Text the agent writes into a rich-text editor, such as a reply in a mail or ticket editor, carries a hidden AI mark where the editor keeps it, and where the page shows a title for it, hovering over the text shows AI-generated. In a plain text box no mark stays in the text; it is only recorded. Your organisation keeps a signed record of each text, so Check content finds it later. Leave the marks in place when you go on editing.

The browser agent's panel after it wrote a reply into a ticket's editor: the Run fill_field? card answered Allowed, Ran fill_field, and the two paragraphs it added

What it never does

  • It never reads the content of a site you have not allowed. A tab on such a site reaches the model only as a tab whose site is not allowed, without its title and address; the site's host name reaches the model once the agent opened it or went to it. On such a site the bar says The agent cannot read …; select Allow on this site (or All sites) to let it. A quick action on text you selected sends that text, even from a site you have not allowed.
  • It never reads out or fills in password fields, nor card fields the site marks as such. A payment field the site does not mark is read and filled like any other field.
  • It never presses a button, such as Send, or Enter in a field without asking you, every time: no earlier answer covers it. Only in Bypass, where your organisation allows it and you switched it on, does what the grant covers run without asking.
  • It never goes to another site in your tab by itself. It is told never to buy, pay, create accounts or accept terms for you; that is an instruction to the model, not a technical block.
  • It is told never to take instructions from the page or a picture: what they say is something to read, not an order to follow. What a page, a PDF, a transcript or a site's tool contains reaches the model marked as the page's text, in a way the page cannot fake, and text that tries to instruct assistants is pointed out to the model; the agent then tells you that the page tries to instruct it. A page can still mislead a model, so read what a card asks before you allow it.
  • It never reaches files on your computer that you did not add.

Permissions it asks for

At installation, Chrome lets the extension open its side panel, keep its settings and your sign-in, add its right-click items, read the titles and addresses of your tabs, search with Chrome's search engine when you switch web search on, and read and act on pages, but only on sites you allow. When you use the selection shortcut or Ask about a screenshot of this page, Chrome lets it into that tab that once. It asks separately to reach your Fadenstack server (when you select Connect) and for each site you allow. You can take any site back in Chrome's settings for the extension.

Where your data goes

  • What you ask, and what the agent reads to answer (the page, tabs, PDF pages, pictures you add), goes to your organisation's server and to the model your administrator chose for the agent.
  • With web search switched on, a search you allow goes to the search engine set in Chrome.
  • A video's captions are fetched from the video's own site, as its player does. With tools that sites offer switched on, what a site's tool is given goes to that site.
  • PDFs are read inside your browser. Pictures are made smaller in your browser before they go. CSV and Markdown files you save stay on your computer.
  • Your conversations stay in this browser, encrypted, each signed-in person's on their own. The key stays in the browser and cannot be read out of it. Your saved prompts stay in this browser too. The events the extension sends to the server say which tools ran and how long they took, not what they read. What the tools read still reaches the server, in the requests to the AI model, as the first point says.
  • Your sign-in is kept in this browser and, unlike the conversations, is not encrypted there. On a computer others use, sign out when you are done: select the gear (Settings) at the top of the panel, then Connection, then Sign out. To remove your conversations and saved prompts as well, select Your data in this browser, then Remove my data….

More on what is kept and who sees it: Your data and privacy.

What it cannot read

What What happens
A scan (pages that are pictures) It says so. With a model that sees pictures, add a screenshot instead
A PDF with a password It says so: open it in a PDF reader instead
A PDF over 50 MB It says so
A PDF on your computer opened in Chrome Add it with Add a PDF from this computer instead
A selection inside Chrome's PDF viewer Copy the passage into your message
A video without captions or a transcript It says so
A part of the page from a site you have not allowed (a frame) It says so. Allow that site in Chrome's settings for the extension (Site access)
Chrome's own pages and the Chrome Web Store Chrome does not let extensions read them

When it does not work

You see What to do
… does not answer. Check the server's address, and that this computer trusts its certificate. The address is wrong, or your computer does not trust the server's certificate. Check the address, or tell your IT department.
There is no agent '…' on … The Agent name is wrong. Select the gear (Settings), then Connection, and connect again with the name your administrator gave you.
You have no access to '…'. Ask your administrator to grant it to you or your team. You are not yet allowed to use this agent.
'…' has no published version yet. or '…' is retired. The agent is not available. Ask your administrator.
Your organisation turned this agent off. Your administrator stopped the agent for now. Your conversations stay in this browser; once it is on again, ask again.
Your account is deactivated. Ask your administrator.
Fadenstack needs to reach the server. Allow it when Chrome asks. You declined Chrome's question. Select Connect again and allow it.
This agent's model does not see pictures. Pictures cannot go to this agent. Ask your administrator if you need them.
Select some text on the page first. A quick action on a selection found none. Select the text, then choose it again.