Chrome Control MCP

Let AI agents use the Chrome you already use -- signed in as you, in tabs of their own, visibly, and without ever taking your screen.

This extension installs the Chrome Control MCP server and registers it with VS Code, so its 54 tools are available to chat without editing an mcp.json or knowing a path. The same server works with any MCP client -- Claude Code, Codex, Cursor, Claude Desktop -- and several of them can share one browser at once.

Three agents from two MCP clients in one Chrome window, each with its own coloured tab group and matching page frame

What it does

Capability	Included
Many agents at once	Several editors, and several agents inside one, each in its own tabs and coloured tab group
Your tabs stay yours	A tab you open never joins an AI group; drag a tab in to hand it over, out to take it back
Works beside you	Drives background tabs and never brings them to the front; an agent's popup goes behind your tab
Always visible	A frame, a named badge and a cursor in the agent's colour on every tab it holds
Cleans up after itself	`browser_release`, or the editor exiting, closes what the agent opened; borrowed tabs are handed back
Uploads without dialogs	File inputs, script-created inputs, drop zones and `showOpenFilePicker`, filled directly in background tabs
A clipboard each	Copy, cut and paste use the agent's own clipboard, never yours or another agent's; pastes carry text, formatting and images into any editor
Semantic control	Accessibility snapshots with element refs bound to the tab they came from
Real input	Click, type, keys, hover, drag, select, scroll, dialogs and media
Browser state	Cookies, storage, permissions, downloads, print, HTTP auth, emulation, and tab recording to `.webm`
In-place updates	The server updates itself without overwriting the copy that is running

The server talks to your ordinary Chrome through native messaging and the Chrome DevTools Protocol. It opens no network listener, and nothing leaves your machine.

Setup

Installing this extension does the first two steps for you:

The server is installed into a per-user directory on first activation. Nothing needs administrator rights.
VS Code is told about it through the MCP server definition provider API, so no configuration file needs editing.
Chrome has to be pointed at the extension folder once, by you. Run Chrome Control MCP: Show the Chrome extension folder to load from the command palette: it copies the folder path and opens it. Then in Chrome open chrome://extensions, turn on Developer mode, choose Load unpacked, and pick that folder.

Or, instead of the folder, install Chrome Control MCP from the Chrome Web Store: one click, and the store keeps it up to date. Google reviews every version first, so the store can trail a release by a few days. Use one or the other in a Chrome profile, not both.

That third step cannot be automated, and this extension does not pretend otherwise. Branded Chrome ignores --load-extension, and a store install is your own click too. It is a one-time step: updates reuse the same folder, and the store updates its copy itself.

Using it with several agents

Each agent should open its own tab with browser_new_tab -- optionally with a short agent name, which gives its tabs their own coloured group -- and pass the returned tab_id to every call. The server tells every client this itself, through the standard MCP instructions. When an agent is done it calls browser_release: tabs it opened close, tabs it borrowed are handed back.

Updating

Marketplace updates and the server's own updater (browser_update_check, then browser_update_apply) converge on the same install, and from 1.8.0 on an older copy never rolls a newer one back. Each version is installed into its own directory behind a stable link, so the editor's server command, Chrome's extension folder and the native-messaging registration never change.

After an update, reload VS Code and press reload on the extension at chrome://extensions so both run the new build.

Commands

Command	What it does
Chrome Control MCP: Repair install	Reinstalls and re-registers the server and native host
Chrome Control MCP: Show the Chrome extension folder...	Copies and opens the folder to load into `chrome://extensions`

Security

The server pins the extension origin, authenticates the same-user, same-executable relay, limits and validates every IPC message, refuses stale element refs and screenshot coordinates, and exposes strict JSON Schemas. Windows uses a current-user ACL-protected named pipe; Linux and macOS use a mode-0600 Unix socket in a private directory with peer credentials.

For inspection-only use, set these in the server environment:

CHROME_CONTROL_MCP_SECURITY_PROFILE=read_only
CHROME_CONTROL_MCP_REDACT_SENSITIVE_OUTPUT=true

The read-only profile exposes 15 tools and independently refuses hidden mutating calls.

Requirements

Google Chrome 116 or later
Windows x64, Linux x64, or macOS (Apple silicon or Intel)

Builds are intentionally unsigned and unnotarized; this project has no signing key and does not need one.

Links

Repository · Tool reference · Security · Verification

Randy Northrup