Tools

The list never changes as you browse, so your agent's context stays flat. The typed path is the default and the cheapest; the visual and low-level tools are there for tasks that need them, and their descriptions tell your agent when.

How your agent gets them

• Loaded up front: browser_navigate, browser_invoke, browser_act, browser_find, browser_read, browser_ask and browser_screenshot. Clients that defer MCP tools behind a search, such as Claude Code, load these seven straight away, because nearly every task needs them and each search costs a turn.
• Found by search: the other 20 carry search hints, so one search finds them.
• Read-only: the tools that only read (browser_capabilities, browser_read, browser_ask, browser_find, browser_inspect, browser_extract, browser_console, browser_network, browser_screenshot, browser_look, browser_downloads and task_status) are marked read-only for your client.
• Hidden when switched off: tools you switch off on the setup page disappear from your agent's list at once (Settings).

Every tool that works on a tab takes an optional tabId; without one, it uses the active tab.

Act through typed tools

browser_capabilities

The full contract of every open tab: each form as a typed tool with its JSON Schema, what its fields currently hold, and the buttons and links with refs. Long pages return a window of 25 links and the true total.

browser_invoke

Calls a page tool that browser_capabilities listed, with arguments matching its schema. It fills, and where allowed submits, the real form in your logged-in browser. A form that changes something asks first (Before it submits).

delegate_goal

Hands Deputy a goal in plain language. Deputy's planner chooses a tool on the tab and its arguments, and calls it. Argument: goal. See Delegating a goal.

task_status

Returns a delegated task by taskId: its state, steps, and result or error.

answer_task

Resumes a task that stopped for a person, a confirmation or a sign-in now done in the browser. Arguments: taskId, answer.

Read without looking

browser_ask

Answers a question about a tab in a sentence; the page never enters your agent's context. Without a working model, returns the page's passages most relevant to the question instead. See Reading pages.

browser_read

The page as plain text. maxChars caps it: 6,000 by default, up to 40,000.

browser_extract

A table or a repeated list as JSON rows, or the page's links. kind: auto (default), table, list or links; limit rows (default 50); scope with ref or selector.

browser_find

Finds elements by text, role or selector, and returns refs with their label, state and value. Reaches inside open and closed shadow roots and off-screen elements. Also exact, includeHidden, and limit (default 20).

browser_downloads

Files the browser downloaded. action: list shows recent downloads, wait waits for one just started by a click (timeoutMs, default 30000), and read returns a finished download by id (the newest without one): text files as text, PDFs as extracted text, others as their path, type and size.

Act on anything

browser_act

Does one thing to an element, by ref or selector. action: click, double_click, right_click, hover, focus, type, press, check, uncheck, select, clear or scroll_into_view.

Typing is sent key by key, so autocompletes react. The result reports any dialog or error the action caused.

browser_scroll

Scrolls the page or an element: direction (down, up), by pixels, or to top or bottom. untilStable keeps scrolling while an infinite feed keeps loading (maxRounds, default 10) and reports how much it grew. It finds the real scroll container in app layouts where the window itself doesn't scroll.

browser_drag

Drags one element onto another: fromRef or fromSelector, toRef or toSelector. Works with HTML5 drag and drop and with pointer-driven libraries such as kanban boards.

browser_upload

Attaches local files to a file input, or drops them onto a drop zone. paths are absolute paths, 25 MB in total.

browser_dialog

Sets how alert, confirm and prompt are answered on a tab your agent is driving. By default they're dismissed (a confirm returns false) and reported in the action result. Call it before the action that opens the dialog: accept, text for a prompt, and persist to keep answering that way.

browser_fill

Sets one field by ref to a value, for inputs Deputy couldn't put into a typed tool. A checkbox takes "true" or "false"; a <select> takes an option's value or visible text.

browser_click

Clicks a button or link that browser_capabilities listed, by ref.

Navigate

browser_navigate

Opens a url and waits for it to load, or moves through history with action: back, forward or reload. The answer lists the new page's tools with their schemas, the start of its text and its actions. With no web page active, it opens a tab of its own rather than using yours.

browser_tabs

action: list every tab, open a url in a new active tab, select a tab by tabId (other tools act on the active tab by default), or close one.

browser_wait

Waits until text appears, textGone disappears, or an element matching selector is visible, for pages that load content after the fact. With no condition, waits timeoutMs (default 10000, at most 60000). Fails with a clear message on timeout.

See and debug

browser_screenshot

Returns a screenshot to your agent, roughly 1,000 to 1,600 tokens per image. fullPage, ref or selector, format (jpeg or png), maxDimension (default 1280). See Seeing and debugging.

browser_look

Deputy Pro. A vision model answers a question about what's visible, or describes the page, in text. Needs the setup page's screenshot switch. See browser_look.

browser_inspect

With a ref or selector: the element's box model, computed CSS (or the properties you name), layout, contrast, overflow and what covers it. With no target: a layout audit of the page.

browser_console

Console messages and page errors. levels (error, warn, info, log, debug), limit, and clear to empty the log after reading.

browser_network

Requests the page made. filter by URL, failedOnly, includeBodies (the first 4 KB of text and JSON bodies), limit, clear.

browser_resize

Sets the viewport's width and height in CSS pixels, at least 200 each, by resizing your real window. reset: true puts the window back.

browser_eval

Runs code in the page and returns its JSON-serialisable result. world: main (default) or isolated. Needs Chrome's Allow User Scripts switch for Deputy.