Browser tools
7. What the agent can do in the browser
The agent gets 18 tools. Some are read-only and never ask for permission; the others change something and go through your approval (depending on the mode).
| Tool | What it does | Permission |
|---|---|---|
browser_tabs_context | Lists the tabs in the conversation's group | Read |
browser_tabs_create | Opens a URL in the group (in the background) | Read |
browser_tabs_attach | Brings an already open tab into the group | Asks |
browser_tabs_close_tab | Closes a tab in the group | Asks |
browser_navigate | Navigates, goes back or forward in the history | Asks |
browser_read_page | Reads the page and returns the elements with references | Read |
browser_find | Finds elements by text, role or name | Read |
browser_get_page_text | Extracts the main text, without HTML | Read |
browser_form_input | Fills in a field | Asks |
browser_computer | Clicks, types, presses keys, scrolls, moves the mouse, takes a screenshot, zooms, waits | Asks |
browser_javascript_tool | Runs JavaScript on the page | Asks |
browser_read_console_messages | Reads the tab's console and errors | Read |
browser_read_network_requests | Reads the tab's network requests | Read |
browser_resize_window | Resizes the window to test layouts | No request |
browser_update_plan | Shows domains and steps for you to approve beforehand | Plan card |
browser_upload_image | Sends files to a file field or drop area | Asks |
browser_gif_creator | Records action frames and exports a GIF | Export asks |
browser_batch | Runs several tools in sequence (up to 20) | Each action asks for its own |
Tool cards
Each call appears in the chat with a label — “Navigate”, “Read page”, “Click on the page”, “Fill field”, “Console”, “Network” — and a status: ok, error or running.
You choose how much detail you want in Settings → Tools in chat:
| Mode | How it looks |
|---|---|
| Grouped (open in a modal) (default) | One summary row per run (“4 tools · view”); the details open in the “Tools of this run” modal |
| Compact rows, one per call | One row per tool, with the error summarized on the row itself |
| Full cards, with parameters and result | Everything open in the chat, including the input JSON and the output |
| Hide tools and reasoning (only “working”) | The chat shows only the activity (“working: reading the page”) |
In grouped mode, the modal has an “Errors only” filter and closes with Esc or by clicking outside.
Screenshots appear as an image inside the card (up to 3 at a time) and are also saved to disk: the file path comes in the agent's reply.