Skip to main content
Glama
Sign In
open-browser-control

open-browser-control

by smankoo
OverviewSchemaSourceRelated ServersDiscussionsAdmin
TypeScript
Local

Open Browser Control

Give AI agents control of your browser — Chrome or Firefox. Works with Kiro, Claude Code, Claude Desktop, Cursor, and any MCP client.

The AI uses your real browser — your cookies, sessions, and logins. When it hits something it can't handle (sign-in, CAPTCHA, MFA), it asks you to step in, then continues where it left off.

Install either the Chrome or the Firefox extension — the MCP server speaks the same protocol to both.

Quick Start

Step 1: Add MCP config

Add to your MCP client's config:

Kiro — add to .kiro/settings/mcp.json (workspace) or ~/.kiro/settings/mcp.json (global):

{
  "mcpServers": {
    "browser": {
      "command": "npx",
      "args": ["-y", "open-browser-control"]
    }
  }
}
claude mcp add browser -- npx -y open-browser-control

Add to your MCP config file:

{
  "mcpServers": {
    "browser": {
      "command": "npx",
      "args": ["-y", "open-browser-control"]
    }
  }
}

Step 2: Install the browser extension

Pick whichever browser you'd like the AI to drive.

Chrome — install from the Chrome Web Store:

Open Browser Control on the Chrome Web Store

Unpack the extension and load it in Firefox:

npx -y open-browser-control --extension firefox

This prints a path like ~/open-browser-control-extension-firefox. Then:

  1. Open about:debugging#/runtime/this-firefox

  2. Click Load Temporary Add-on…

  3. Pick the manifest.json inside that folder

Firefox clears unsigned temporary add-ons on restart. For a permanent install, use Firefox Developer Edition (xpinstall.signatures.required=false) or sign the extension via AMO.

The extension is auto-installed to ~/open-browser-control-extension/ on first run. Load it in Chrome:

  1. Open chrome://extensions/

  2. Enable Developer mode (toggle in top right)

  3. Click Load unpacked

  4. Select the folder: ~/open-browser-control-extension/

Done. The extension auto-connects when your agent starts. No servers to run, no buttons to click.

Related MCP server: Wbrowser

How It Works

AI Agent          MCP Server              Browser Extension
(Kiro,       ◄──► (npx open-browser-  ◄──► (Chrome: CDP, or
 Claude, ..)      control)                  Firefox: scripting API)
              stdio    ws://localhost:9334    auto-connect

  1. Your agent starts the MCP server automatically (from the config above)

  2. MCP server starts a WebSocket bridge on localhost:9334

  3. The browser extension (Chrome or Firefox) auto-connects (polls every 2s until it finds the bridge)

  4. Agent sends tool calls → extension executes (via Chrome DevTools Protocol on Chrome, or the WebExtensions scripting API on Firefox) → results flow back

Both extensions implement the same JSON protocol over WebSocket, so the MCP server, bridge, and agents never need to know which browser is on the other end.

User/AI Handoff

Mode

What happens

Collaborative (default)

Both user and AI interact with the page

AI Control

AI drives, user watches

User Control

AI paused, user takes over

AI browsing → hits login page → calls browser_request_user("Please sign in")
  → user signs in → clicks "Done" in side panel → AI continues, now authenticated

Browser Tools

19 tools available to the AI:

Tool

What it does

browser_navigate

Go to a URL

browser_get_dom

Get interactive elements with positions and text (primary way to read pages)

browser_get_page_info

URL, title, dimensions, scroll position

browser_execute_js

Run JavaScript in page context

browser_click

Click by selector, text, or coordinates

browser_type

Type text, optionally clear first or press Enter

browser_scroll

Scroll up/down/left/right

browser_keypress

Press any key with modifiers

browser_hover

Hover over an element

browser_select_option

Pick from a dropdown

browser_wait

Wait for element, text, or fixed time

browser_screenshot

Capture page as PNG (use sparingly — DOM tools are faster)

browser_request_user

Ask user to take over (sign in, CAPTCHA, etc.)

browser_new_tab_group

Create a named tab group for a task

browser_new_tab

Open a new tab in the current group

browser_close_tab

Close a tab

browser_switch_tab

Switch to a tab by ID

browser_list_tabs

List all open tabs in this session

browser_set_session_name

Set the session name (shown on tab group)

Standalone Use (no MCP)

If you're not using an MCP client:

npx -y open-browser-control --bridge    # starts WebSocket bridge only

Connect your agent to ws://localhost:9334 and send JSON messages:

{"type": "action", "action": "navigate", "id": "1", "params": {"url": "https://example.com"}}
{"type": "action", "action": "click", "id": "2", "params": {"text": "Sign In"}}
{"type": "action", "action": "get_dom", "id": "3"}

CLI

npx -y open-browser-control                 # Start MCP server (default)
npx -y open-browser-control --bridge        # Standalone WebSocket bridge
npx -y open-browser-control --extension     # Print extension install path
npx -y open-browser-control --port 9000     # Custom port
npx -y open-browser-control --help          # Help

Development

git clone https://github.com/smankoo/open-browser-control
cd open-browser-control
npm install
npm run build    # builds both browsers → dist/chrome, dist/firefox
                 # and packages to extension/ and extension-firefox/
npm run dev      # watch mode
npm start        # run MCP server locally

Requirements

  • Chrome 116+ or Firefox 128+

  • Node.js 18+

Available Tools

18 tools
browser_clickC

Click on a page element. Specify by CSS selector, visible text, or x/y coordinates.

Inputs: x, y, text, button, +1 more.

browser_execute_jsA

Execute JavaScript in the page context and return the result. Useful for extracting specific data, reading computed styles, or interacting with page APIs — often more precise than screenshots for getting exact values.

Inputs: expression (required).

browser_get_domA

Get the page DOM structure. Returns interactive elements with their text, roles, and positions. This is the primary way to understand what is on the page — use it to find elements to click, read text content, and understand page layout. Faster and more detailed than screenshots.

Inputs: selector, simplified.

browser_get_page_infoA

Get current page metadata: URL, title, dimensions, scroll position. Use this to quickly check where you are and whether navigation succeeded — much faster than a screenshot.

Inputs: no named parameters.

browser_hoverC

Move mouse over an element to trigger hover effects.

Inputs: x, y, text, selector.

browser_keypressC

Press a keyboard key, optionally with modifiers.

Inputs: key (required), modifiers.

browser_list_tabsA

List all open browser tabs in this session.

Inputs: no named parameters.

browser_navigateB

Navigate to a URL in the current tab.

Inputs: url (required).

browser_new_tabB

Open a new browser tab.

Inputs: url.

browser_new_tab_groupA

Create a new named tab group for a task. Use this to organize browsing by task — each group gets a label visible in Chrome. The new group becomes the active group; new tabs and navigations happen inside it.

Inputs: name (required), url.

browser_request_userA

Ask the user to perform an action (e.g., sign in, solve CAPTCHA). AI pauses until user signals done.

Inputs: message (required), timeout.

browser_screenshotA

Take a screenshot of the current browser page. Returns the file path to a saved PNG image. Screenshots are slow and heavy — prefer browser_get_dom and browser_get_page_info to understand page content and structure. Use screenshots only when you need to verify visual layout, debug rendering issues, or when DOM inspection is insufficient (e.g. canvas, images, charts). The response includes a pageRect describing which page-coordinate region the image depicts. To zoom into fine detail (especially after a full-page shot), call again with rect:{x,y,width,height} narrowed to a sub-region of that pageRect — coordinates are page CSS pixels.

Inputs: rect, fullPage, selector.

browser_scrollB

Scroll the page in a direction.

Inputs: direction (required), amount, selector.

browser_select_optionB

Select an option from a dropdown element.

Inputs: selector (required), text, value.

browser_set_session_nameA

Set a descriptive name for this browser session. The name appears on the Chrome tab group so the user can identify what task this session is working on. Call this early with a short task summary (e.g. "Search flights to Paris", "Debug login page").

Inputs: name (required).

browser_switch_tabA

Switch to a specific browser tab by its ID.

Inputs: tabId (required).

browser_typeC

Type text into the focused element or a specific element.

Inputs: text (required), clear, selector, pressEnter.

browser_waitB

Wait for a condition: element to appear, text to show, or fixed time.

Inputs: ms, text, timeout, selector.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Deploy Server
    One endpoint for every MCP toolRayrun runs every MCP server behind one URL and holds the credentials. Each call is decided per client, tool, and argument, then recorded.ray.runAd

    TDQS

    A3.5/5.0

    Scored across 18 tools

    Disambiguation5/5

    Each tool targets a distinct action: clicking, typing, scrolling, DOM inspection, screenshot, tab management, JS execution, user requests, etc. Descriptions clearly differentiate between DOM-based and screenshot-based understanding, and between navigation and interaction primitives.

    Naming Consistency5/5

    All tools follow the consistent snake_case pattern with the verb_noun structure and browser_ prefix. There is no mixing of conventions, making names predictable and easy to infer.

    Tool Count5/5

    18 tools is well-scoped for a browser control server. It covers navigation, interaction, DOM reading, tab management, and special actions like user requests, without being bloated or too sparse.

    Completeness2/5

    Obvious gaps exist: no tool to close a tab, no back/forward navigation, and no cookie or storage management. These are standard browser automation operations whose absence will cause agent failures.

    Maintenance

    ActivityInactive
    ResponsivenessNo recent community issues

    Resources

    Related MCP Connectors

    Related MCP Servers

    • Wbrowser

      matveikarpov64
      A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI assistants and terminal users to control a logged-in Chrome browser, performing actions like opening pages, searching, filling forms, and taking screenshots without re-authentication.
      MIT

    • browser-control

      kierandrewett
      F
      license
      Not graded
      quality
      B
      maintenance
      Enables AI agents to control existing Chrome and Firefox browser sessions locally, including tab management, navigation, form filling, screenshots, and persistent scripts, without remote relays or debugging ports.
      9 npm
      -