> ## Documentation Index
> Fetch the complete documentation index at: https://help.casaconect.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Browser sessions

> When a task needs a real website, such as a registry, a portal or a form, the agent can drive a browser in the cloud while you watch, and hand it to you when it needs a person.

Some information does not come as a file or a search result. It is behind a portal, a form or a map viewer. For that, the agent can open a **browser session**: a real browser running in the cloud that it navigates the way a person would.

## Web search or a browser session?

|                    | Web search                                                             | Browser session                                                          |
| :----------------- | :--------------------------------------------------------------------- | :----------------------------------------------------------------------- |
| **Good for**       | Public facts: a regulation, an address, a company's registered details | Websites that must be *used*: search forms, portals, multi-step look-ups |
| **Speed and cost** | Fast, negligible                                                       | Slower. Billed per second of browser time.                               |
| **Needs you**      | Never                                                                  | Sometimes, to sign in                                                    |

The agent chooses web search when that is enough. You can also ask for a browser explicitly: *"open the Catastro site and look up this reference."*

## What you see

A browser session appears in the conversation as a card with its state:

| State                      | Meaning                                                                       |
| :------------------------- | :---------------------------------------------------------------------------- |
| **Working in the browser** | The agent is navigating. Choose **Watch live** to see the page as it sees it. |
| **Waiting for you**        | The site needs a person. See below.                                           |
| **Session finished**       | Done. Choose **Watch replay** to see a recording of what happened.            |
| **Session failed**         | The session ended with an error. The agent reports what it managed to do.     |

Each action is described as it happens: *"Inspected the page"*, *"Captured a screenshot"*, and so on.

## When the browser needs you

Signing in, solving a CAPTCHA or confirming a one-time code are jobs for a person. When the agent reaches one, it stops and shows **The browser needs you**.

<Steps>
  <Step title="Open the browser tab">
    You take control of the same live browser the agent was using.
  </Step>

  <Step title="Do what only you can do">
    Sign in, complete the challenge or enter the code. You type your password into the website yourself. The agent does not handle it, and it is not written into the conversation.

    A sign-in can be kept for later sessions as a saved login, so that you are not asked again every time. Sign out on the website if you do not want it remembered.
  </Step>

  <Step title="Choose Done — continue">
    The agent picks up where it stopped, now signed in. Choose **Cancel** to end the session instead.
  </Step>
</Steps>

## Safety

A browser can reach the whole internet and can act with your logins, so it is the most tightly controlled thing the agent does.

<AccordionGroup>
  <Accordion title="Attended only" icon="eye">
    Browser sessions exist only in conversations where you are present. They are never available to work started by email, to matter runs with nobody watching, or to helpers. A message from outside cannot cause a browser to open.
  </Accordion>

  <Accordion title="Driving the browser is a sensitive action" icon="shield-check">
    Opening a session, navigating and acting on a page are **sensitive** under [permission levels](/agent/permission-levels). At the default level they pass the safety review. At *Ask for approval* they ask you. Reading the page and taking screenshots count as reads.
  </Accordion>

  <Accordion title="Web pages are data, not instructions" icon="shield-alert">
    Text on a web page is handled like any other outside content. The agent reads it and does not obey it. See [Agent safety](/security/agent-safety).
  </Accordion>

  <Accordion title="Recorded" icon="video">
    Sessions are recorded so that you can review exactly what was done. Be aware that a recording shows whatever was on screen, including pages viewed after you signed in.
  </Accordion>

  <Accordion title="A separate provider" icon="server">
    Cloud browsers are run by a specialist provider listed under [Subprocessors](/security/subprocessors). The provider's own model handles page-level actions such as finding the right button. The conversation model plans the visit and reads the results.
  </Accordion>
</AccordionGroup>

## Good uses

* Looking up a cadastral reference and saving the descriptive certificate to the project
* Checking a company's entry in a public register
* Downloading a document from a portal you have access to, after you sign in
* Filling in a public form up to the point of submission, for you to review and submit

<Warning>
  Ask the agent to stop **before** anything irreversible, such as submitting a filing, making a payment or accepting terms, and do that step yourself. The agent is instructed to be cautious, but the responsibility for a submission is yours.
</Warning>

## Cost

Browser time is billed per second and appears under **Browsing the web** in **Settings → Access & billing → Usage**. Sessions are closed when the task is done, you can ask the agent to close one at any time, and a session that is left open ends by itself after 30 minutes.