Browser

What browser MCP server does

The browser MCP server gives an MCP-compatible assistant access to a real Chrome, Edge, or Brave browser. It can open a URL, wait for the page to load, and return page content from the browser's rendered state. This makes it suitable for sites that depend on JavaScript, require an authenticated session, or use SSO, CAPTCHA, or anti-bot controls.

The server keeps tabs organized by hostname. When a page from the same domain is requested again, it can reuse the existing tab and its session state. A tab can also be closed explicitly to clear that domain's browser state and force a new session on the next visit.

How it works

The assistant communicates with the server over MCP, while the server controls the browser through the Chrome DevTools Protocol. A typical workflow is to load a page, enter values into one or more fields, click a visible control, and then read the updated DOM. The browser_get_current_html operation reads the current page without navigating or reloading, which is useful after an interaction.

JavaScript can be run in the active page context when selectors or protocol-level actions are not enough. The result includes execution metadata such as timing, truncation status, and whether the URL changed. Screenshots provide a PNG view of the loaded page for charts, images, popups, complex layouts, or other content that is difficult to interpret from HTML.

Setup and configuration

The browser MCP server requires Chrome, Edge, or Brave and Node.js 18 or newer. The repository documents a standard configuration that runs the npm package with npx:

{
  "mcpServers": {
    "mcpbrowser": {
      "command": "npx",
      "args": ["-y", "mcpbrowser@latest"]
    }
  }
}

Add this configuration through the MCP client you use. The project also documents installation paths for clients including Claude Desktop, Claude Code, Cline, Codex, Cursor, Gemini CLI, LM Studio, VS Code, and Windsurf.

Tools and capabilities

The available tools cover navigation, inspection, interaction, and session cleanup:

• browser_fetch_webpage opens a URL, reuses a same-domain tab when possible, waits for loading, and returns content.
• browser_execute_javascript runs a script in the active page and returns its result with metadata.
• browser_click_element targets an element by CSS selector or visible text, scrolls it into view, and waits for page stability.
• browser_type_text fills one or more input fields and clears existing values by default.
• browser_get_current_html extracts the current DOM without another navigation.
• browser_scroll_page scrolls an already-loaded page.
• browser_take_screenshot captures the loaded page as a PNG.
• browser_close_tab removes the tab for a hostname and resets its next visit.

Limitations and notes

The browser MCP server exposes webpage content to the connected AI assistant, which may send that content to its configured language-model provider. Avoid using it with sensitive or private pages unless the agent and provider are trusted. The project identifies this combination of private-data access, untrusted content, and external communication as a security risk and documents tool annotations for client safety controls.

A local compatible browser and Node.js installation are required. Screenshot capture is intended for an already-loaded page, while current HTML, scrolling, and element interaction also operate on an existing page rather than providing a separate HTTP-fetching interface.