MCP 浏览器代理

MCP Browser Agent

Features

• Advanced Browser Automation

  • Navigate to any URL with customizable load strategies
  • Capture full-page or element-specific screenshots
  • Perform precise DOM interactions (click, fill, select, hover)
  • Execute arbitrary JavaScript in browser context with console logs capture

• Powerful API Client

  • Execute HTTP requests (GET, POST, PUT, PATCH, DELETE)
  • Configure request headers and body content
  • Process response data with JSON formatting
  • Error handling with detailed feedback

• MCP Resource Management

  • Access browser console logs as resources
  • Retrieve screenshots through MCP resource interface
  • Persistent session with headful browser instance

• AI Agent Capabilities

  • Chain multiple browser operations for complex tasks
  • Follow multi-step instructions with intelligent error recovery
  • Technical task automation through natural language instructions

Demo

Timestamps:

Click on any timestamp to jump to that section of the video

00:00 - Google Search for MCP
Navigation to Google homepage and search for "Model Context Protocol". Demonstration of Claude Desktop using the MCP integration to perform a basic web search and process the results.

00:33 - Screenshot Capture
Taking a screenshot of the search results with a custom filename and showcasing it in Finder. Shows how Claude can capture and save visual content from web pages during browser automation.

01:00 - Wikipedia Search
Navigation to Wikipedia.org and search for "Model Context Protocol". Illustrates Claude's ability to interact with different websites and their search functionality through the MCP integration.

01:38 - Dropdown Menu Interaction I
Navigation to a test website (the-internet.herokuapp.com/dropdown) and selection of "Option 1" from a dropdown menu. Demonstrates Claude's capability to interact with form elements and make selections.

01:56 - Dropdown Menu Interaction II
Changing the selection to "Option 2" from the same dropdown menu. Shows Claude's ability to manipulate the same form element multiple times and make different selections.

02:09 - Login Form Completion
Navigation to a login page (the-internet.herokuapp.com/login) and filling in the username field with "tomsmith" and password field with "SuperSecretPassword!". Demonstrates form filling automation.

02:28 - Login Submission
Submitting the login credentials and completing the authentication process. Shows Claude's ability to trigger form submissions and navigate through multi-step processes.

02:36 - API Request Execution
Performing a GET request to JSONPlaceholder API endpoint. Demonstrates Claude's capability to make direct API calls and process the returned data through the MCP integration.

Requirements

• Node.js 20 or higher
• Claude Desktop
• A browser: Google Chrome, or one of Playwright's builds (see below)

Browser Support

The package supports the following browsers:

• Chrome (default)
• Firefox
• Microsoft Edge
• WebKit (Safari engine)

Chrome (the default) uses your installed Google Chrome. If Chrome is not installed, the agent falls back to Playwright's own Chromium build and logs that it did.

Playwright's browsers are not downloaded on first use — install the ones you want once:

npx playwright install chromium   # fallback when Google Chrome is absent
npx playwright install firefox
npx playwright install webkit
npx playwright install msedge

A clone followed by npm install downloads Chromium, Firefox and WebKit automatically through the bundled @playwright/browser-* packages. Installing from Smithery or an MCPB bundle does not, so run the command above for any browser other than Google Chrome.

macOS 13 (Ventura): Playwright is pinned to 1.61, the last release that ships browsers for macOS 13; Playwright 1.62 dropped it. Google Chrome keeps working on any version.

Note about Safari: Playwright doesn't provide direct support for Safari browser. Instead, it uses WebKit, which is the browser engine that powers Safari.

Note about Edge: Selecting edge launches your installed Microsoft Edge (before 0.10.0 it silently opened plain Chromium). If Edge is not installed, the agent falls back to Playwright's Chromium. Technically, in Playwright, Edge is launched using the Chromium browser instance with the 'msedge' channel parameter because Microsoft Edge is based on Chromium.

Installation

Installing via Smithery

Install the packaged bundle from the Smithery server page, or from the CLI:

npx -y @smithery/cli@latest mcp add imprvhub/mcp-browser-agent --client claude

Installing Manually

1. Clone or download this repository:

git clone https://github.com/imprvhub/mcp-browser-agent
cd mcp-browser-agent

2. Install dependencies:

npm install

3. Build the project:

npm run build

Running the MCP Server

There are two ways to run the MCP server:

Option 1: Running manually

1. Open a terminal or command prompt
2. Navigate to the project directory
3. Run the server directly:

node dist/index.js

Keep this terminal window open while using Claude Desktop. The server will run until you close the terminal.

Option 2: Auto-starting with Claude Desktop (recommended for regular use)

The Claude Desktop can automatically start the MCP server when needed. To set this up:

Configuration

The Claude Desktop configuration file is located at:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

Edit this file to add the Browser Agent MCP configuration. If the file doesn't exist, create it:

{
  "mcpServers": {
    "browserAgent": {
      "command": "node",
      "args": ["ABSOLUTE_PATH_TO_DIRECTORY/mcp-browser-agent/dist/index.js",
      "--browser",
      "chrome"
    ]
    }
  }
}

Important: Replace ABSOLUTE_PATH_TO_DIRECTORY with the complete absolute path where you installed the MCP

• macOS/Linux example: /Users/username/mcp-browser-agent
• Windows example: C:\\Users\\username\\mcp-browser-agent

If you already have other MCPs configured, simply add the "browserAgent" section inside the "mcpServers" object. Here's an example of a configuration with multiple MCPs:

{
  "mcpServers": {
    "otherMcp1": {
      "command": "...",
      "args": ["..."]
    },
    "otherMcp2": {
      "command": "...",
      "args": ["..."]
    },
    "browserAgent": {
      "command": "node",
      "args": [
        "ABSOLUTE_PATH_TO_DIRECTORY/mcp-browser-agent/dist/index.js",
      "--browser",
      "chrome"
    ]
    }
  }
}

Browser Selection

The MCP Browser Agent supports multiple browser types. By default, it uses Chrome, but you can specify a different browser in several ways:

Option 1: Configuration File

Create or edit the file .mcp_browser_agent_config.json in your home directory:

{
  "browserType": "chrome"
}

Supported values for browserType are:

• chrome - Uses installed Chrome (default)
• firefox - Uses Firefox 'Nightly' browser
• webkit - Uses WebKit engine (Note: This is not Safari itself but the WebKit rendering engine that powers Safari)
• edge - Uses Microsoft Edge

Note about Safari: Playwright doesn't provide direct support for Safari browser. Instead, it uses WebKit, which is the browser engine that powers Safari. The WebKit implementation in Playwright provides similar functionality but is not identical to the Safari browser experience.

Option 2: Command Line Argument

When starting the MCP server manually, you can specify the browser type:

node dist/index.js --browser firefox

Option 3: Environment Variable

Set the MCP_BROWSER_TYPE environment variable:

MCP_BROWSER_TYPE=firefox node dist/index.js

Other environment variables:

Option 4: Claude Desktop Configuration

When configuring the MCP in Claude Desktop's claude_desktop_config.json, you can specify the browser type:

{
  "mcpServers": {
    "browserAgent": {
      "command": "node",
      "args": [
        "ABSOLUTE_PATH_TO_DIRECTORY/mcp-browser-agent/dist/index.js",
        "--browser",
        "chrome"
      ]
    }
  }
}

Technical Implementation

MCP Browser Agent is built on the Model Context Protocol, enabling Claude to interact with a headful browser through Playwright. The implementation consists of four main components:

1. Server (index.ts)

   • Initializes the MCP server with Model Context Protocol standard protocol
   • Configures server capabilities for tools and resources
   • Establishes communication with Claude through the stdio transport

2. Tools Registry (tools.ts)

   • Defines browser and API tool schemas
   • Specifies parameters, validation rules, and descriptions
   • Registers tools with the MCP server for Claude's discovery

3. Request Handlers (handlers.ts)

   • Manages MCP protocol requests for tools and resources
   • Exposes browser logs and screenshots as queryable resources
   • Routes tool execution requests to the appropriate handlers

4. Executor (executor.ts)

   • Manages browser and API client lifecycle
   • Implements browser automation functions using Playwright
   • Handles API requests with proper error handling and response parsing
   • Maintains stateful browser session between commands

Agent Capabilities

Unlike basic integrations, MCP Browser Agent functions as a true AI agent by:

• Maintaining persistent browser state across multiple commands
• Capturing detailed console logs for debugging
• Storing screenshots for reference and review
• Managing complex interaction sequences
• Providing detailed error information for recovery
• Supporting chained operations for complex workflows

Available Tools

Browser Tools

API Tools

Resource Access

The MCP Browser Agent exposes the following resources:

• browser://logs - Access browser console logs
• screenshot://[name] - Access screenshots by name

Example Usage

Here are some realistic examples of how to use the MCP Browser Agent with Claude:

Basic Browser Navigation

Navigate to the Google homepage at https://www.google.com

Take a screenshot of the current page and name it "google-homepage"

Type "weather forecast" in the search box

Simple Interactions

Navigate to https://www.wikipedia.org and search for "Model Context Protocol"

Go to https://the-internet.herokuapp.com/dropdown and select the option "Option 1" from the dropdown

Basic Form Filling

Navigate to https://the-internet.herokuapp.com/login and fill in the username field with "tomsmith" and the password field with "SuperSecretPassword!"

Go to https://the-internet.herokuapp.com/login, fill in the username and password fields, then click the login button

Simple JavaScript Execution

Go to https://example.com and execute a JavaScript script to return the page title

Navigate to https://www.google.com and execute a JavaScript script to count the number of links on the page

Basic API Requests

Perform a GET request to https://jsonplaceholder.typicode.com/todos/1

Make a POST request to https://jsonplaceholder.typicode.com/posts with appropriate JSON data

These examples represent the actual capabilities of the MCP Browser Agent and are more realistic about what it can accomplish in its current state.

Troubleshooting

"Server disconnected" error

If you see the error "MCP Browser Agent: Server disconnected" in Claude Desktop:

1. Verify the server is running:

   • Open a terminal and manually run node dist/index.js from the project directory
   • If the server starts successfully, use Claude while keeping this terminal open

2. Check your configuration:

   • Ensure the absolute path in claude_desktop_config.json is correct for your system
   • Double-check that you've used double backslashes (\\) for Windows paths
   • Verify you're using the complete path from the root of your filesystem

Tools return nothing

Versions before 0.9.0 returned every result in a legacy toolResult field alongside an empty content array. Current MCP clients read only content, so each tool appeared to succeed with no output. Update to 0.9.0 or later.

Navigation fails with ERR_NAME_NOT_RESOLVED while API calls work

Playwright's browser is a separate executable from Node. A per-application firewall such as Little Snitch can let Node through while blocking the newly downloaded browser. Allow the browser binary under ~/Library/Caches/ms-playwright/ (macOS) in your firewall.

Browser not appearing

If the browser doesn't launch or you don't see it:

1. Check if the specified browser is installed

   • Verify that you have the browser (Chrome, Firefox, Edge, or Safari/WebKit) installed on your system
   • The browser drivers are handled automatically by Playwright

2. Restart the server and Claude Desktop

   • Kill any existing node processes that might be running the server
   • Restart Claude Desktop to establish a fresh connection

Browser process not closing properly

Since 0.10.0 the agent handles the two causes seen in practice:

• The client quits (you close Claude Desktop): the server now notices its input closing, shuts the browser down and exits. Before, the open browser kept both processes running.
• You quit the browser window yourself (Cmd+Q, or Quit from the dock): on macOS the window closes but Playwright keeps the process alive. The agent now closes the browser from its side when its page closes, and the next tool call starts a fresh one.

If a process still lingers after a crash or a forced kill:

1. Manually close the browser process:

   • Windows: Press Ctrl+Shift+Esc to open Task Manager, find the Chrome/Chromium process and end it
   • macOS: Open Activity Monitor (Applications > Utilities > Activity Monitor), find the Chrome/Chromium process and click the X to terminate it
   • Linux: Run ps aux | grep chrome or ps aux | grep chromium to find the process, then kill <PID> to terminate it

2. Note about browser compatibility:

   • This issue has been observed primarily with Chromium and Chrome
   • Firefox and Playwright's built-in browser don't typically experience this problem

[!CAUTION] This MCP integration is built on Playwright, which has known issues and bugs that may affect its operation. Please report any issues you encounter with the browser automation to Playwright's GitHub issues. The Playwright team is continuously working to address these issues, but this agent provides a foundation for browser automation capabilities with Claude Desktop despite these limitations.

Development

Project Structure

• src/index.ts: Main entry point and MCP server initialization
• src/tools.ts: Tool schemas and registration
• src/handlers.ts: MCP request handlers for tools and resources
• src/executor.ts: Tool implementation logic using Playwright

Building

npm run build

Watching for Changes

npm run watch

Testing

npm run build
npm test

The tests start the built server and talk to it over stdio the way an MCP client does. They check that tool output arrives where clients read it, that file: URLs are refused, and a full headless browser session against a local fixture page (navigate, evaluate, fill, screenshot). The browser test is skipped when no Playwright browser is installed.

Security Considerations

Pages the agent visits can contain text written to steer the model — prompt injection. The defaults below narrow what such a page can make the agent do:

• file: URLs are refused unless MCP_BROWSER_ALLOW_FILE_URLS=true. Otherwise a page could have the model open a local file (an SSH key, a credentials file) and pass it to an api_post call. The check runs before a browser is launched.
• Screenshot names are reduced to a plain file name. The name used to be joined into the save path unmodified, so ../ in it wrote outside the chosen folder and created any directories along the way.
• Output is capped at 100,000 characters per response body or script result, so one large response cannot flood the conversation.

The api_* tools can reach any address the machine can, localhost included — intended for testing local APIs, but worth knowing when the agent browses untrusted sites.

[!IMPORTANT] This MCP integration provides Claude with autonomous browser control capabilities. Please review our Security Policy for important information about prohibited uses, security implications, and best practices.

The MCP Browser Agent is designed for legitimate automation tasks but could potentially be misused. Users are responsible for ensuring their use complies with all applicable laws, terms of service, and ethical guidelines. See our detailed Security Policy for more information.

Contributing

Contributions to the MCP Browser Agent are welcome! Here are some areas where you can help:

• Adding new browser automation capabilities
• Improving error handling and recovery
• Enhancing screenshot and resource management
• Creating useful workflows and examples
• Optimizing performance for complex operations

License

This project is licensed under the Mozilla Public License 2.0 - see the LICENSE file for details.

Related Links

• Model Context Protocol
• Claude Desktop
• Playwright Documentation
• MCP Series