Browser
Read rendered web pages, inspect images, and act on a page that stays open.
Read a public web page, including pages that need JavaScript to display their content. Results include readable text, the page title, final URL, HTTP status, links, and images. Inspect images when information appears in charts, diagrams, screenshots, photos, or logos. An agent can also keep a page open for a conversation and click, type, choose options, and press keys on it, for tasks such as searching a site or filling in a public form.
This built-in capability is managed by Yekar.AI. You do not need to connect a separate account. It is available when enabled for your workspace.
Tools
| Tool | Access |
|---|---|
| Read a web page | Read |
| Read an image | Read |
| Open a page to act on | Write |
| Read the open page | Read |
| Click on the open page | Write |
| Type into the open page | Write |
| Choose an option on the open page | Write |
| Press a key on the open page | Write |
| Close the open page | Write |
Usage
- Provide an absolute web page URL.
- Use the returned links and nearby text to find company websites, related pages, documents, and contact details. Follow actual links instead of guessing addresses.
- Read an image accepts a direct image URL, or a page URL and an image index from a page read. Describe visible details and mark anything unreadable or uncertain as a gap.
- HTTP errors such as 404 mean the requested source was not retrieved. Try another source or report the gap; an error does not prove the information is absent.
- To let an agent act on pages, reference Act on a web page in its Procedure. Agents that only reference Read a web page can read pages but cannot click, type, or submit.
- A click or the Enter key can submit a form, place an order, or send a message. To have a person confirm these steps first, add the Browser's tools to the agent's Approvals. The request shows the name of the button, link, or field, and the step is refused if the page has changed so that the name no longer matches.
- Each conversation has its own open page. It keeps cookies only until it is closed, and it closes on its own after about 15 minutes without use, so a step that waits longer than that for a person's answer may need the page opened again.
- Pages are visited as an anonymous visitor. The Browser does not sign in and does not fill password fields, so private sites and content behind a login are unavailable.
- Controls inside embedded frames are not available to act on, and at most 200 controls are listed for one page.
- The browser scrolls to load content that appears as it comes into view. A bounded scan may leave content unloaded. Hidden sections, such as tabs, are returned with their links; open a relevant section to read its text and images.
- Long text and image lists are paginated. Continue with the returned nextOffset or nextImageOffset to read more. Each read reloads the page, so changing pages may produce different contents or image indices; refresh the inventory if needed. Link lists may also be shortened. Treat unread content as a gap, not evidence that information is absent.