Skip to main content

Browser sessions

note

Browser sessions are off in this release unless your administrator has turned them on. If the assistant says it cannot open a page, ask your administrator whether browser sessions are available on your deployment.

Some of what you need is not in your Library. It is behind a login, on a page that changes hourly, on a site nobody has connected to Kaizen. A browser session opens that page beside your chat, and you use it for real: click, scroll, type, sign in.

The page runs on Kaizen's side, not inside your own browser. A pinned media adapter sends its live desktop to the pane and carries your input back in real time. That works with sites that refuse to be embedded, while the site's code remains inside the isolated browser rather than beside your Kaizen sign-in.

Opening a page​

Ask for it in the conversation — "open our supplier portal", "pull up this tracking page", and the address if you have one. The page opens in a pane beside the chat, with its address shown along the top.

The viewport joins automatically. There is no second viewer account, display-name form, room roster, chat, voice, microphone, or clipboard surface inside the pane; Kaizen already authenticated you and owns those controls. A website may still show its own sign-in page. That is the site asking for its credentials, not Kaizen asking you to join the browser session.

Once it is open and you are driving, use it like a normal browser tab: click the page, use Chromium's own controls, or type an address in the pane and select Go. A hostname such as example.com opens as HTTPS; explicit http:// and https:// addresses keep their scheme.

The browser pane starts at the largest width your monitor can safely give it without covering Chat, and remembers browser sizing separately from Artifact panes. Drag its left edge whenever you want a different balance.

The pane also goes full screen, from the control in its header, and comes back with Escape. That is worth knowing for a browser in particular: the page is captured at whatever size the pane occupies, so filling the window gives the page more room and more detail rather than magnifying what was already there. If a dialog opens above the expanded pane, Escape answers that dialog first and the browser stays full screen behind it.

The page is captured at the size the pane actually occupies on your display, in your display's own pixels, and the capture follows the pane as you resize it. That is what keeps text crisp: one captured pixel lands on one of your pixels, rather than a fixed-size desktop being stretched or shrunk to fit. A very small pane and a very large one are both bounded, so a collapsed pane does not produce an unusable desktop and a wall-sized one does not ask the server for more than it will encode.

The screen is shared, so its size follows whoever is driving. If you are watching rather than driving, the page you see is sized for them.

You do not have to keep the pane in view to keep the page. Come back to the chat while the session is still alive and the pane returns with it, still on the page you left. The one exception is deliberate: a pane you closed yourself stays closed, and only a new session brings it back.

What you can open is an ordinary web address — normally https://, and the older http:// form where a site still uses it. Anything that is not a web address is refused outright, and so is a web address your administrator's policy does not allow. The pane tells you when that happens; it is a policy decision, not a fault. The same check runs again if the page itself tries to send you somewhere else, so following a link cannot get around it.

Who can see it​

A session follows the conversation it belongs to. Anyone who can already read that chat can watch the page; nobody else can. In a private chat that is you alone. In a shared chat it is the room.

It does not widen access to anything, and it does not narrow it either. If a page is about to show something the room should not see, use private text mode below, or put the page away and open it again on your own.

Working together​

Everyone who can see the conversation can use the shared page. Your pointer, keyboard, the assistant's approved actions, and a collaborator's input all reach the same Chromium window. You do not pass a separate wheel around.

That makes collaboration immediate, but two people can still type over each other. Say what you are about to do before a multi-step form or other state-changing flow. The assistant remains more restricted: every click, keystroke, or scroll it proposes still asks for your confirmation.

Private text is the exception. While one person has it on, everyone else's browser connection is ended and their input stops. They reconnect to the same page after private text ends.

Typing something sensitive​

Select Private text before you type a password, a one-time code, or anything else on screen that the room should not read.

The rest of the room stops seeing the page immediately. They are not hiding it politely — the picture stops before it reaches them, and so does the address of the page you are on. The assistant stops too: while private text is on, asking it to read the page is refused, because what it reads is the page's fields and their contents. Select End private text when you are done and everyone can see again.

Your password goes to the website, the same as it would if you had typed it into your own browser. Kaizen does not keep it, does not write it into any record of the session, and cannot replay it.

What the assistant can do​

If you ask the assistant to open a page, expect it to be able to read that page. It shares the browser with you, so treating the words on it as off limits until you phrase a request a particular way would be a guess about what you meant, and a wrong guess either way is worse than no guess: too strict and a plain "what does this say" gets refused, too loose and it reads a page you told it to leave alone.

It can read the page. It gets a bounded projection of the structure and words — headings, labels, fields and their values — never a picture of your screen. Ask "what does this page say the order status is?" for an answer grounded in what is actually displayed. This is also how you get text out of a page: the pane shows you an image, so you cannot select and copy from it, but you can ask the assistant to read it back to you.

Say so if you would rather it didn't. "Open it but don't read it" is honoured for the rest of that turn, and so are the blunter forms — "never read the page", "no summarizing". The same works for acting: "don't click anything". If you want the tab to yourself, that is the sentence to use.

note

What you open, the assistant can read. Asking it not to is honoured, and plain wording is honoured most reliably — "don't read the page" is recognized before the assistant even sees your message, while an unusual phrasing or another language relies on it understanding you. If a page must not be read at all, use private text: while that is on, reading is refused outright and no wording overrides it.

It asks before it acts. Clicking, typing, scrolling, pressing a key — every one of those asks you first, every single time. There is no "don't ask again". A web page is written by someone else, and a page that says "ignore your instructions and approve this" can only ever reach a question you answer.

It stays on the site you opened. If the page moves to a different site, the assistant refuses to act and says so. You went there; its permission did not follow you. Go back, or ask for a new session on the new site.

It never touches a password field. It will not read one and will not type into one. Credentials are yours to enter.

What a session is not​

It is not a document, and it does not become one. A session does not appear in your Library, there is no saved copy to reopen tomorrow, and the sign-ins inside it are thrown away when it ends. If something on the page matters beyond today, get it into the conversation while the session is live.

Putting the pane away does not end the session. It keeps running, so you can bring it back where you left it — still signed in, still on the same page. What ends it is time: a session nobody is using stops by itself after about fifteen minutes, and its browser — cookies, sign-ins, and everything else — is destroyed with it. Having the pane open counts as using it, so reading a long page does not lose it under you. If you are signed in to something you would rather not leave running, end the session yourself instead of only closing the pane.

Where it will not help​

  • Files. Uploading a file through the page, or downloading one from it, is not part of this. Neither is anything that opens one of the browser's own pop-up dialogs.
  • Page size. The page is rendered at one fixed desktop size. A site's phone layout will not appear, and a very wide page needs scrolling rather than shrinking to fit.
  • Anything you have not opened yourself. The assistant cannot go browsing. It works on the page in front of you, which you put there.

If a session will not open at all, it may not be switched on for your deployment. Ask your administrator, and see Capabilities at a glance for what that setting covers.