MCP Server Browser

What agent-infra/mcp-server-browser does

agent-infra/mcp-server-browser gives MCP clients access to browser automation backed by Puppeteer. It can launch or use a local browser, connect to an existing browser through a Chrome DevTools Protocol endpoint or WebSocket endpoint, and expose the resulting session to an LLM-powered client.

The default interaction model uses accessibility snapshots and Puppeteer's label index rather than pixel coordinates. This keeps browser state represented as structured data. An optional vision mode adds screenshot-based interaction for pages where accessibility information alone is insufficient.

The package is intended for browser automation workflows in the Browser Automation category. One documented MCP tool is browser_navigate, which accepts a URL and navigates the browser to it.

How it works

For local use, an MCP client starts the package as a Node.js stdio process. The server can also listen on a configured port and provide either an SSE endpoint or a Streamable HTTP endpoint. Streamable HTTP is the recommended remote transport in the command examples.

The server can use Chrome, Edge, or Firefox, and it can run in headed or headless mode. Browser selection, executable location, persistent user data, proxy settings, viewport size, and user-agent behavior are controlled through command-line arguments. Existing browser sessions can be reached with CDP or WebSocket connection endpoints.

JavaScript and TypeScript applications can import createServer and connect to it with an in-memory MCP transport. This avoids requiring the command-line interface to be installed by the application’s end user.

Setup and configuration

Node.js 18 or newer is required. A typical stdio configuration runs the package with npx:

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

For remote operation, start the package with --port, then configure the client with either /sse or /mcp. The server binds to localhost by default; use --host to change the binding address. A DISPLAY environment setting can be supplied when running with VNC or another virtual display.

Useful options include --browser, --cdp-endpoint, --ws-endpoint, --executable-path, --headless, --proxy-server, --proxy-bypass, --user-agent, --user-data-dir, --viewport-size, and --output-dir. Runtime HTTP headers can override viewport size, vision coordinate factors, and user-agent values for a request. Header names are case-insensitive.

Tools and capabilities

• Navigates to a supplied URL with the browser_navigate MCP tool.
• Uses structured accessibility data by default.
• Enables optional screenshot-based vision mode with --vision.
• Connects to local browsers or remote CDP and WebSocket browser endpoints.
• Supports Chrome, Edge, and Firefox selection.
• Provides stdio, SSE, Streamable HTTP, and in-memory MCP connection patterns.
• Allows per-runtime viewport, coordinate-factor, and user-agent headers.