BrowserMCP

MCP server that gives AI assistants full control over your browser. Works with Claude Code, Cursor, and any MCP-compatible client.

The key advantage: uses your real browser session with all cookies, auth tokens, and extensions — no headless browser, no 403 errors.

Architecture

AI Client (Claude Code)  --stdio-->  MCP Server (Node.js)  --WebSocket-->  Chrome Extension
                                     port 12800

Three components:

MCP Server — Node.js process, communicates with AI via stdio, bridges to browser via WebSocket
WebSocket Bridge — runs on port 12800, connects server to extension
Chrome Extension — Manifest V3, executes commands in the browser using Chrome APIs

36 Tools

Core

| Tool | Description | |------|-------------| | browser_navigate | Open URL and return page content | | browser_read_page | Read content from active or specific tab (text/html/full) | | browser_list_tabs | List all open tabs with IDs, URLs, titles | | browser_close_tab | Close a tab by ID | | browser_click | Click element by CSS selector | | browser_type | Type text into input field | | browser_screenshot | Capture visible tab as PNG | | browser_execute_js | Run JavaScript on the page |

Navigation

| Tool | Description | |------|-------------| | browser_scroll | Scroll by direction/pixels or to CSS selector | | browser_back | Navigate back in history | | browser_forward | Navigate forward in history | | browser_switch_tab | Activate and focus a tab | | browser_keyboard | Press keys with modifiers, inserts characters into inputs | | browser_hover | Hover over element (mouseenter/mouseover) | | browser_select | Select option in dropdown |

Data Extraction

| Tool | Description | |------|-------------| | browser_get_links | Extract all links, optionally filtered | | browser_get_elements | Query elements by CSS selector with attributes | | browser_extract_table | Parse table into structured JSON (headers + rows) | | browser_extract_meta | Extract meta tags, OpenGraph, Twitter Cards, JSON-LD | | browser_extract_images | Get all images with src, alt, dimensions | | browser_get_cookies | Read cookies for a URL | | browser_get_storage | Read localStorage or sessionStorage |

Utilities

| Tool | Description | |------|-------------| | browser_wait_for | Wait for element to appear (MutationObserver) | | browser_new_tab | Open new tab | | browser_reload | Reload tab (with optional cache bypass) | | browser_set_storage | Write to localStorage or sessionStorage | | browser_set_cookies | Set a cookie | | browser_find_text | Search text on page with context |

Monitoring

| Tool | Description | |------|-------------| | browser_console_log | Capture console.log/warn/error (start/get/stop) | | browser_network_log | Capture fetch/XHR requests (start/get/stop) | | browser_get_computed_style | Read computed CSS properties | | browser_inject_css | Inject custom CSS into page | | browser_highlight | Highlight elements with colored outline | | browser_readability | Extract main article content (Reader Mode) | | browser_watch_changes | Watch DOM mutations (start/get/stop) |

Window Management

| Tool | Description | |------|-------------| | browser_new_window | Open new window (with size, incognito) | | browser_close_window | Close window by ID | | browser_resize_window | Resize window | | browser_move_tab | Move tab between windows | | browser_pin_tab | Pin/unpin tab | | browser_mute_tab | Mute/unmute tab |

Advanced Interaction

| Tool | Description | |------|-------------| | browser_right_click | Right-click (contextmenu event) | | browser_double_click | Double-click | | browser_drag_drop | Drag and drop between elements |

Installation

1. Clone and build

git clone https://github.com/Smotree/BrowserMCP.git
cd BrowserMCP
npm install
npm run build

2. Load Chrome extension

Open chrome://extensions/
Enable Developer mode (top right)
Click Load unpacked
Select the extension/ folder

3. Configure your MCP client

Auto-install (recommended)

Run the install script — it auto-detects Claude Code, VS Code, Cursor, and Claude Desktop:

npm run install-mcp        # install to all detected clients
npm run install-mcp -- --uninstall  # remove from all

Manual setup

Claude Code — ~/.claude.json (user-scoped, all projects)

Via CLI:

claude mcp add --transport stdio --scope user browser -- node /path/to/BrowserMCP/server/dist/index.js

Or manually add to ~/.claude.json:

{
  "mcpServers": {
    "browser": {
      "type": "stdio",
      "command": "node",
      "args": ["/path/to/BrowserMCP/server/dist/index.js"]
    }
  }
}

For a single project only, add .mcp.json to the project root with the same format.

VS Code — %APPDATA%/Code/User/mcp.json (Win) / ~/Library/Application Support/Code/User/mcp.json (Mac)

{
  "servers": {
    "browser": {
      "command": "node",
      "args": ["/path/to/BrowserMCP/server/dist/index.js"],
      "type": "stdio"
    }
  }
}

If the file already has other servers, just add "browser": { ... } inside "servers".

Cursor — %APPDATA%/Cursor/User/mcp.json (Win) / ~/Library/Application Support/Cursor/User/mcp.json (Mac)

Same format as VS Code above.

Claude Desktop — %APPDATA%/Claude/claude_desktop_config.json (Win) / ~/Library/Application Support/Claude/claude_desktop_config.json (Mac)

{
  "mcpServers": {
    "browser": {
      "command": "node",
      "args": ["/path/to/BrowserMCP/server/dist/index.js"]
    }
  }
}

4. Verify

The extension badge shows ON (green) when connected, OFF (red) when disconnected. The extension auto-reconnects within 1-2 seconds when the server starts.

Testing

A test page and two test scripts are included:

# Serve the test page
npm run test:serve

# Quick test — verifies all tools with DOM state checks
# (close Claude Code first to free port 12800)
npm run test:quick

# Demo test — visual walkthrough of all tools for screen recording
npm run test:demo

Both test scripts auto-kill any process on port 12800 before starting.

Configuration

| Environment Variable | Default | Description | |---------------------|---------|-------------| | BROWSER_MCP_WS_PORT | 12800 | WebSocket port for server-extension communication |

Multi-instance (HTTP Relay)

When multiple IDE windows are open, each launches its own BrowserMCP server. The first instance owns port 12800 and connects to the Chrome extension directly. Subsequent instances detect the port is busy and automatically switch to HTTP relay mode — they proxy commands through the first instance's /exec endpoint. This is transparent: all tools work the same regardless of which instance you use.

Project Structure

BrowserMCP/
  extension/              Chrome extension (Manifest V3)
    background.js           Service worker — WebSocket client, command handlers
    manifest.json           Extension manifest
    popup.html/js           Connection status popup
    icons/                  Extension icons
  server/                 MCP server (TypeScript)
    src/
      index.ts              Entry point — stdio transport + WS bridge
      mcp-server.ts         Creates MCP server, registers tools
      ws-bridge.ts          WebSocket server with request correlation
      types.ts              Shared types
      tools/                Tool definitions by category
        core.ts             navigate, read_page, list_tabs, etc.
        navigation.ts       scroll, back, forward, keyboard, etc.
        extraction.ts       get_links, extract_table, get_cookies, etc.
        utilities.ts        wait_for, new_tab, reload, find_text, etc.
        interaction.ts      right_click, double_click, drag_drop
        monitoring.ts       console_log, network_log, inject_css, etc.
        window.ts           new_window, resize_window, pin_tab, etc.
        content.ts          highlight, readability, watch_changes, etc.
  test/
    test-page.html          Test page exercising all 36 tools
    quick-test.mjs          Automated test with assertions
    demo-test.mjs           Visual demo for screen recording

How It Works

MCP client (Claude Code) spawns node server/dist/index.js as a child process
Server communicates with the client via stdin/stdout (MCP protocol)
Server starts a WebSocket server on port 12800
Chrome extension connects to the WebSocket and waits for commands
When a tool is called, server sends a command to the extension via WebSocket
Extension executes the command using Chrome APIs and returns the result
Server forwards the result back to the MCP client

License

MIT

MCP Servers