Agent Chrome Tool Reference: Every MCP Tool and Parameter

Agent Chrome exposes two tool sets from the same server. Coding agents get the agent set, which matches Claude in Chrome and includes computer and javascript_tool. Chat apps such as Claude.ai and ChatGPT get the directory set, which has one tool per action (click, type_text, scroll and so on).

Tool names follow Claude in Chrome, so prompts and skills written for it work unchanged. Every tool also accepts an optional browser argument (a profile id or name from list_browsers) and most accept an optional tabId.

Common parameters

Browsers and pairing

These tools run on the server and are in both tool sets.

{ "name": "select_browser", "arguments": { "browser": "Work" } }

Tabs and groups

tab_groups actions are list, create, update, ungroup, move and close. Colors are grey, blue, red, yellow, green, pink, purple, cyan and orange.

{ "name": "tab_groups", "arguments": { "action": "update", "title": "QA run", "color": "green" } }

Navigation

{ "name": "navigate", "arguments": { "url": "https://news.ycombinator.com" } }

Reading the page

{ "name": "read_page", "arguments": { "filter": "interactive" } }
{ "name": "find", "arguments": { "query": "Sign in button" } }

Mouse, keyboard and screenshots: computer (agent set)

computer is one tool with many actions. Coordinates use the pixel space of the latest screenshot, so take a screenshot first, or use a ref.

hover and mouse_move move the pointer without clicking. Native HTML5 drag and drop works with left_click_drag.

{ "name": "computer", "arguments": { "action": "left_click", "ref": "ref_12" } }
{ "name": "computer", "arguments": { "action": "type", "text": "agent chrome" } }
{ "name": "computer", "arguments": { "action": "key", "text": "Enter" } }
{ "name": "computer", "arguments": { "action": "left_click_drag", "start_ref": "ref_4", "ref": "ref_9" } }

Forms, files and dialogs

{ "name": "form_input", "arguments": { "ref": "ref_7", "value": "kai@example.com" } }
{ "name": "file_upload", "arguments": { "ref": "ref_21", "paths": ["/Users/me/report.pdf"] } }
{ "name": "wait_for", "arguments": { "text": "Order confirmed", "timeoutMs": 15000 } }

Debugging

{ "name": "javascript_tool", "arguments": { "text": "document.title" } }
{ "name": "read_console_messages", "arguments": { "onlyErrors": true } }
{ "name": "read_network_requests", "arguments": { "urlPattern": "/api/", "includeBodies": true } }

Recording

See recording for details.

{ "name": "record", "arguments": { "action": "start", "cursor": "natural", "showClicks": true } }
{ "name": "record", "arguments": { "action": "stop" } }

Window and history

Chat-app (directory) tool set

Claude.ai and ChatGPT get these tools instead of computer, javascript_tool, history_search and tab_groups. They also get every tool above that is not marked "agent set only": the browser tools, tabs, navigate, the reading tools, form_input, file_upload, upload_image, handle_dialog, wait_for, record, gif_creator, resize_window and screenshot.

{ "name": "click", "arguments": { "ref": "ref_12" } }
{ "name": "type_text", "arguments": { "ref": "ref_3", "text": "hello" } }
{ "name": "press_key", "arguments": { "keys": "Enter" } }

Read-only and destructive hints

Every tool carries annotations so chat apps can ask before risky actions. These are read-only: list_tab_groups, list_browsers, tabs_context, screenshot, read_page, find, get_page_text, read_console_messages, read_network_requests, wait_for, history_search, scroll, hover, wait and zoom.

These are marked destructive because their effects may not be undoable: tabs_close, tab_groups, close_tab_group, click, press_key, computer, javascript_tool, file_upload and upload_image. Every tool that talks to a browser is marked open-world.

Related pages

• Docs overview
• Use Agent Chrome from Claude and ChatGPT
• Record screenshots, MP4 clips and GIFs
• Debug console and network errors
• Fill forms and upload files

Frequently asked questions

What is the difference between the agent and directory tool sets?

The agent set is for coding agents and includes the general computer and javascript_tool tools. The directory set is for chat apps in the Claude and ChatGPT directories and splits actions into separate tools such as click and type_text, each with honest read-only and destructive hints.

Are the tool names the same as Claude in Chrome?

Yes. Where Claude in Chrome has a counterpart, Agent Chrome uses the same name, so prompts and skills written for it work unchanged.

Should I use a ref or coordinates?

Use a ref when you have one. Refs come from read_page or find and survive layout changes better. Coordinates use the pixel space of the latest screenshot, so take a screenshot first.

How do I choose between two Chrome profiles?

Pass browser: "Work" on any call, or call select_browser once for the session. list_browsers shows each profile and whether it is online. See multiple Chrome profiles.

Can an agent read my browsing history?

Only if you turn on history access in the extension popup. It is off by default, and history_search is available to coding agents only.

Updated 2026-11-04 · Agent Chrome is open source (MIT) and operated by Circo, Inc.