Browser MCP

Browser Use MCP Server

A fast, lightweight Model Context Protocol (MCP) server that empowers LLMs with browser automation via Puppeteer’s structured accessibility data, featuring optional vision mode for complex visual understanding and flexible, cross-platform configuration.

Key Features

• ⚡ Fast & lightweight. Utilizes Puppeteer's label index, not pixel-based input and accessibility DOM tree.

• 👁️ Vision Mode Support. Optional visual understanding capabilities for complex layouts and visual elements when structured data isn't sufficient.

• 🤖 LLM-optimized. No vision models needed, operates purely on structured data, less context reducing context token usage.

• 🧩 Flexible Runtime Configuration. Customize viewport size, coordinate system factors, and User-Agent at runtime via HTTP headers.

• 🌐 Cross-Platform & Extensible. Support for remote and local browsers, the use of a custom browser engine.

Requirements

• Node.js 18 or newer

• VS Code, Cursor, Windsurf, Claude Desktop or any other MCP client

Getting started

Local (Stdio)

First, install the Browser MCP server with your client. A typical configuration looks like this:

{
 "mcpServers": {
 "browser": {
 "command": "npx",
 "args": [
 "@agent-infra/mcp-server-browser@latest"
 ]
 }
 }
}

Install in VS Code You can also install the Browser MCP server using the VS Code CLI:

# For VS Code
code --add-mcp '{"name":"browser","command":"npx","args":["@agent-infra/mcp-server-browser@latest"]}'

After installation, the Browser MCP server will be available for use with your GitHub Copilot agent in VS Code.

Install in Cursor Go to Cursor Settings -> MCP -> Add new MCP Server. Name to your liking, use command type with the command npx @agent-infra/mcp-server-browser. You can also verify config or add command like arguments via clicking Edit.

{
 "mcpServers": {
 "browser": {
 "command": "npx",
 "args": [
 "@agent-infra/mcp-server-browser@latest"
 ]
 }
 }
}

Install in Windsurf Follow Windsuff MCP documentation. Use following configuration:

{
 "mcpServers": {
 "browser": {
 "command": "npx",
 "args": [
 "@agent-infra/mcp-server-browser@latest"
 ]
 }
 }
}

Install in Claude Desktop Follow the MCP install guide, use following configuration:

{
 "mcpServers": {
 "browser": {
 "command": "npx",
 "args": [
 "@agent-infra/mcp-server-browser@latest"
 ]
 }
 }
}

Remote (SSE / Streamable HTTP)

At the same time, use --port $your_port arg to start the browser mcp can be converted into SSE and Streamable HTTP Server.

# normal run remote mcp server
npx @agent-infra/mcp-server-browser --port 8089

# run with DISPLAY environment for VNC or other virtual display
DISPLAY=:0 npx @agent-infra/mcp-server-browser --port 8089

You can use one of the two MCP Server remote endpoint:

• Streamable HTTP(Recommended): http://127.0.0.1::8089/mcp

• SSE: http://127.0.0.1::8089/sse

And then in MCP client config, set the url to the SSE endpoint:

{
 "mcpServers": {
 "browser": {
 "url": "http://127.0.0.1::8089/sse"
 }
 }
}

url to the Streamable HTTP:

{
 "mcpServers": {
 "browser": {
 "type": "streamable-http", // If there is MCP Client support
 "url": "http://127.0.0.1::8089/mcp"
 }
 }
}

In-memory call

If your MCP Client is developed based on JavaScript / TypeScript, you can directly use in-process calls to avoid requiring your users to install the command-line interface to use Browser MCP.

import { Client } from '@modelcontextprotocol/sdk/client/index.js';
import { InMemoryTransport } from '@modelcontextprotocol/sdk/inMemory.js';

// type: module project usage
import { createServer } from '@agent-infra/mcp-server-browser';
// commonjs project usage
// const { createServer } = await import('@agent-infra/mcp-server-browser')

const client = new Client(
 {
 name: 'test browser client',
 version: '1.0',
 },
 {
 capabilities: {},
 },
);

const server = createServer();
const [clientTransport, serverTransport] = InMemoryTransport.createLinkedPair();

await Promise.all([
 client.connect(clientTransport),
 server.connect(serverTransport),
]);

// list tools
const result = await client.listTools();
console.log(result);

// call tool
const toolResult = await client.callTool({
 name: 'browser_navigate',
 arguments: {
 url: 'https://www.google.com',
 },
});
console.log(toolResult);

Configuration

Browser MCP server supports following arguments. They can be provided in the JSON configuration above, as a part of the "args" list:

> npx @agent-infra/mcp-server-browser@latest -h
 -V, --version output the version number
 --browser browser or chrome channel to use, possible values: chrome, edge, firefox.
 --cdp-endpoint CDP endpoint to connect to, for example "http://127.0.0.1:9222/json/version"
 --ws-endpoint WebSocket endpoint to connect to, for example "ws://127.0.0.1:9222/devtools/browser/{id}"
 --executable-path path to the browser executable.
 --headless run browser in headless mode, headed by default
 --host host to bind server to. Default is localhost. Use 0.0.0.0 to bind to all interfaces.
 --port port to listen on for SSE and HTTP transport.
 --proxy-bypass comma-separated domains to bypass proxy, for example ".com,chromium.org,.domain.com"
 --proxy-server specify proxy server, for example "http://myproxy:3128" or "socks5://myproxy:8080"
 --user-agent specify user agent string
 --user-data-dir path to the user data directory.
 --viewport-size specify browser viewport size in pixels, for example "1280, 720"
 --output-dir path to the directory for output files
 --vision Run server that uses screenshots (Aria snapshots are used by default)
 -h, --help display help for command

Runtime Configuration

The browser runtime requires configuration for Viewport Size, Vision Model Coordinate Factors, and User Agent. These can be passed through corresponding HTTP headers:

Header Description x-viewport-size Browser viewport size, format: width,height separated by comma x-vision-factors Vision model coordinate system factors, format: x_factor,y_factor separated by comma x-user-agent User Agent string, defaults to system User Agent if not specified

Note: Header names are case-insensitive.

Example:

x-viewport-size: 1920,1080
x-vision-factors: 1.0,1.0
x-user-agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36

Docker

We have unified the deployment of VNC and MCP under a single URL endpoint, The Dockerfile and DockerHub image will be published together! video

API

Tools

Tool Name Description Parameters browser_click Click an element on the page, before using the tool, use browser_get_clickable_elements to get the index of the element, but not call browser_get_clickable_elements multiple times index (number, optional): Index of the element to click browser_close Close the browser when the task is done and the browser is not needed anymore browser_close_tab Close the current tab browser_evaluate Execute JavaScript in the browser console script (string, required): JavaScript code to execute, () => { /* code */ } browser_form_input_fill Fill out an input field, before using the tool, Either 'index' or 'selector' must be provided selector (string, optional): CSS selector for input field, priority use index, if index is not provided, use selector index (number, optional): Index of the element to fill value (string, required): Value to fill clear (boolean, optional): Whether to clear existing text before filling browser_get_clickable_elements Get the clickable or hoverable or selectable elements on the current page, don't call this tool multiple times browser_get_download_list Get the list of downloaded files browser_get_markdown Get the markdown content of the current page browser_get_text Get the text content of the current page browser_go_back Go back to the previous page browser_go_forward Go forward to the next page browser_hover Hover an element on the page, Either 'index' or 'selector' must be provided index (number, optional): Index of the element to hover selector (string, optional): CSS selector for element to hover browser_navigate Navigate to a URL url (string, required): browser_new_tab Open a new tab url (string, required): URL to open in the new tab browser_press_key Press a key on the keyboard key (string, required): Name of the key to press or a character to generate, such as Enter, Tab, Escape, Backspace, Delete, Insert, F1, F2, F3, F4, F5, F6, F7, F8, F9, F10, F11, F12, ArrowLeft, ArrowRight, ArrowUp, ArrowDown, PageUp, PageDown, Home, End, ShiftLeft, ShiftRight, ControlLeft, ControlRight, AltLeft, AltRight, MetaLeft, MetaRight, CapsLock, PrintScreen, ScrollLock, Pause, ContextMenu browser_read_links Get all links on the current page browser_screenshot Take a screenshot of the current page or a specific element name (string, optional): Name for the screenshot selector (string, optional): CSS selector for element to screenshot index (number, optional): index of the element to screenshot width (number, optional): Width in pixels (default: viewport width) height (number, optional): Height in pixels (default: viewport height) fullPage (boolean, optional): Full page screenshot (default: false) highlight (boolean, optional): Highlight the element browser_scroll Scroll the page amount (number, optional): Pixels to scroll (positive for down, negative for up), if the amount is not provided, scroll to the bottom of the page browser_select Select an element on the page with index, Either 'index' or 'selector' must be provided index (number, optional): Index of the element to select selector (string, optional): CSS selector for element to select value (string, required): Value to select browser_switch_tab Switch to a specific tab index (number, required): Tab index to switch to browser_tab_list Get the list of tabs browser_vision_screen_capture Take a screenshot of the current page for vision mode browser_vision_screen_click Click left mouse button on the page with vision and snapshot, before calling this tool, you should call browser_vision_screen_capture first only once, fallback to browser_click if failed factors (array, optional): Vision model coordinate system scaling factors [width_factor, height_factor] for coordinate space normalization. Transformation formula: x = (x_model * screen_width * width_factor) / width_factor y = (y_model * screen_height * height_factor) / height_factor where x_model, y_model are normalized model output coordinates (0-1), screen_width/height are screen dimensions, width_factor/height_factor are quantization factors, If the factors are unknown, leave it blank. Most models do not require this parameter. x (number, required): X pixel coordinate y (number, required): Y pixel coordinate

Resources

Resource Name URI Pattern Description MIME Type Browser console logs console://logs text/plain Browser Downloads download://{name} Automatic identification based on file extension Browser Screenshots screenshot://{name} Automatic identification based on file extension

Development

Access http://127.0.0.1:6274/:

npm run dev

Extended CLI Entry

#!/usr/bin/env node

const {
 BaseLogger,
 setConfig,
 addMiddleware,
} = require('@agent-infra/mcp-server-browser/request-context');

class CustomLogger extends BaseLogger {
 info(...args) {
 console.log('custom');
 console.log(...args);
 }
}

addMiddleware((req, res, next) => {
 console.log('req', req.headers);
 next();
});

setConfig({
 logger: new CustomLogger(),
});

// start server
require('@agent-infra/mcp-server-browser/index');