Seeing and debugging pages
Screenshots when your agent needs pixels, Pro's text descriptions when it doesn't, and the console, network and layout tools for frontend work.
The typed tools keep the page out of your agent's context. Frontend work is the exception: sometimes your agent has to see the page, or read its console. These tools are there for that, and their descriptions tell your agent when a cheaper one would do.
browser_screenshot
Takes a screenshot and returns the image to your agent. An image costs it roughly 1,000 to 1,600 tokens, so it's for when your agent needs to see something: layout and styling, charts, images, canvas.
| Argument | Does |
|---|---|
fullPage | Stitches the whole scrollable page, including anything off the right edge. Costs more |
ref or selector | Captures one element whole, even one wider or taller than the window |
format | jpeg (default, small) or png (sharper text, much larger) |
maxDimension | The longest edge in pixels, 1280 by default |
Chrome only captures the visible tab, so Deputy brings the tab to the front first. A page or element bigger than the window is scrolled and stitched from up to 12 captures.
browser_look
Deputy Pro. A vision model looks at the page and answers your agent in text: it describes the page, or answers a question about what's visible, such as the layout, an image, a chart, or whether something looks broken. It costs your agent a few hundred tokens instead of an image, and works in clients that can't show images, such as Codex.
browser_look sends a screenshot of the tab to Deputy's server, so it needs your permission first: turn on Allow Deputy Pro to send a screenshot on the setup page. It's off by default. What happens to the screenshot is on Privacy.
browser_inspect
Numbers instead of a screenshot, for CSS debugging.
- With a
reforselector: the element's box model, position, computed CSS (or only thepropertiesyou name), flex or grid layout, visibility, text contrast, overflow, and what's covering it. - With no target: a layout audit of the whole page: horizontal overflow and its culprits, clipped text, broken images, missing alt text, small tap targets, low-contrast text, layout shift and fonts.
browser_console and browser_network
browser_console: uncaught exceptions with their stacks, unhandled rejections, failed resources, and console output from when an agent started driving the tab. Filter bylevels. Browser-internal messages, such as CSP and CORS errors, aren't visible without the debugger.browser_network: the requests the page made, with method, URL, status, time, type and size.failedOnlyshows only 4xx and 5xx responses and network errors;includeBodiesadds text and JSON response bodies (the first 4 KB) captured while an agent drives the tab.
browser_resize
Sets the viewport to a width and height in CSS pixels, for example 375 × 812 to check a mobile layout. It resizes your real window, so your agent should call it with reset: true when it's done, which puts the window back. A size bigger than your screen is clamped to the largest that fits.
browser_eval
Runs JavaScript in the page and returns its JSON-serialisable result; promises are awaited. It runs in the page's own world by default, where it sees the page's globals, or with world: "isolated", apart from them. It needs Chrome's Allow User Scripts switch for Deputy (how to turn it on); until then, the error says how.
Switching these off
On the setup page, under What your agent may do, you can switch off seeing pages as images, reading the console and network, uploads, and reading downloads. The extension refuses those operations, and your agent stops seeing the tools at once. See Settings.