What AI Browser Agents Do: The 19 Tools Behind a Click

An AI browser agent turns "fill out this form" into a sequence of small, concrete actions called tools. Prophet gives Claude 19 of them: two to see the page, three to act on elements, four to navigate, three to read and scroll the page, three to wait, and four to manage tabs.

Knowing what each one does helps you write instructions that work the first time. Below is every tool, grouped the way the extension groups them, followed by how Claude chains them and how to prompt for reliable results.

How the Tool Loop Works

When you send a message, Claude reads your request and decides whether it needs the page. If it does, it emits a tool call: a structured request naming one tool and its inputs. The Prophet extension runs that call in your own tab, in your logged-in session, and sends the result back. Claude then calls another tool or answers you.

Almost every interaction starts with a snapshot. The snapshot lists the page's interactive elements from the accessibility tree, each with a short unique ID (UID), for example uid=Ab12Cd3E button "Submit". Later calls target that UID, so "click the Sign Up button" becomes "click uid=Ab12Cd3E". Targeting by UID from the accessibility tree is sturdier than CSS selectors, because labels and roles survive most redesigns.

Observation: Seeing the Page (2 tools)

Take snapshot captures the accessibility tree of the current tab: buttons, links, inputs and other interactive elements with their roles, names, current values and UIDs. Claude calls it before interacting and again after the page changes. On very large pages the snapshot is cut at a size limit, with a note telling Claude to search for anything not listed.

Search snapshot searches the latest snapshot for elements matching a text query (case-insensitive, partial match). It is how Claude finds "the Export button" on a dashboard with hundreds of elements without reading all of them.

Element Interaction: Acting on the Page (3 tools)

Click element clicks an element by its UID: buttons, links, checkboxes, radio buttons, menu items. It can also double-click. If the element is outside the viewport, it is scrolled into view first.

Fill element enters text into an input or textarea by UID. It scrolls the field into view, focuses it, replaces any existing value and types the new one through Chrome's own input system, so the page reacts as if you had typed it. It works with text, email, password, number and search fields, and with rich text editors.

Hover element moves the pointer over an element to trigger hover states, such as menus that open on hover or tooltips that hold the information you need.

There is no separate "select option" or "check box" tool: checkboxes, radio buttons and custom dropdown menus are handled with click, after Claude reads their current state in the snapshot.

Navigation: Moving Between Pages (4 tools)

Navigate loads a URL in the current tab and waits for the page to finish loading before returning.

Go back and go forward work like the browser buttons, for example returning to a results list after checking one item.

Reload page refreshes the current page, which helps when a page is stale or stuck.

Page: Reading and Scrolling (3 tools)

Scroll page scrolls up, down, left or right by a number of pixels (500 by default), or jumps to the very top or bottom. After scrolling, Claude takes a new snapshot to see what appeared.

Read page content returns the page URL, title and the cleaned main text, without HTML. This is the tool behind "summarize this article" or "what does this contract say about termination?", where reading matters more than clicking.

Get page info returns the current URL, title and viewport. It tells Claude where it ended up after a redirect or a multi-step flow.

Waiting: Letting the Page Settle (3 tools)

Modern pages load content after the first render, and acting too early is the most common cause of failed automation.

Wait for selector waits until an element matching a CSS selector appears (optionally until it is visible), up to a timeout of 10 seconds by default.

Wait for navigation waits until the page reports it has finished loading, up to 30 seconds by default.

Wait for timeout pauses for a set number of milliseconds, for animations or slow widgets that give no signal.

Tabs: Working Across Pages (4 tools)

List tabs shows the open tabs so Claude can find the one it needs.

Switch tab moves focus to a tab by its ID, which is how Claude compares information across two sites.

Open new tab opens a URL in a new tab, in the foreground or background, without losing the page you are on.

Close tab closes a tab by its ID once Claude is done with it.

How Claude Chains the Tools

For "find the total on this invoice", Claude needs no interaction: it reads the page content and answers.

For "subscribe to the monthly plan", it takes a snapshot, searches it for the plan, clicks the monthly option, fills any required fields, clicks the confirm button, waits for navigation and takes a fresh snapshot to check the confirmation. If a dialog or a validation error appears, the next snapshot shows it and Claude adjusts.

For "compare the pricing on these two sites", it opens the second site in a new tab, reads both pages, switches between them as needed and writes the comparison.

Instructions that name an element's purpose help most. "Click the Subscribe button in the pricing section" maps directly to a role and a label in the snapshot; "click the blue button on the left" does not, because the snapshot describes what elements are, not what they look like.

Reliability and Error Handling

• Element not found: if a UID no longer exists (the page changed), the tool returns an error and Claude takes a new snapshot and retries with the current UID.
• Element off-screen: click scrolls the element into view before clicking.
• Page loading: navigate, back, forward and reload wait for the page to finish loading; the wait tools cover content that loads later.
• Dynamic content: each snapshot reflects the page as rendered at that moment, including content added by JavaScript, so Claude works from the current state rather than the first load.

Tips for Better Results

• Describe the goal, not the clicks. "Fill out the contact form with name John and email john@example.com" lets Claude plan the snapshot, fills and submit itself.
• Use labels and functions. "The search box", "the submit button", "the price in the first row" match what the snapshot contains.
• Split long workflows. For multi-page tasks, give one phase at a time: "First open the settings page and find the notification preferences."
• Ask for a check. "Did the form submit? What does the confirmation say?" makes Claude take a fresh snapshot or read the page and report what is actually there.

To see the loop in action, visit the how it works page. For why reading the accessibility tree beats screenshots, see what an AI web agent is, and for the tasks people use these tools for, explore the use cases directory.