Chrome DevTools MCP

Chrome DevTools MCP is a Model Context Protocol server that lets an AI coding agent control and inspect a real, running Chrome browser. It is maintained by the Chrome DevTools team at Google in the ChromeDevTools/chrome-devtools-mcp repository.

At its core, the server uses Puppeteer for browser automation and adds genuine DevTools capabilities on top: recording and analysing performance traces, inspecting network requests, reading console messages, and capturing screenshots.

According to the documentation, tools are grouped into these categories:

Input automation (clicks, keyboard input, forms), navigation automation, emulation (e.g. network throttling, device profiles), performance (trace recording and analysis), network (request inspection), and debugging (console, stack traces).

For simpler tasks there is a reduced slim mode (--slim) with a smaller tool set. It requires an LTS version of Node.js and a current stable version of Chrome; other Chromium-based browsers are not officially supported.

The server runs directly via npx without any prior installation:

npx -y chrome-devtools-mcp@latest

In an MCP client, it is registered through the usual configuration file:

{
  "mcpServers": {
    "chrome-devtools": {
      "command": "npx",
      "args": ["-y", "chrome-devtools-mcp@latest"]
    }
  }
}

The official client configuration guide covers setup for a large number of tools, including Claude Code, Cursor, Codex, VS Code/Copilot, Gemini CLI, Windsurf, Cline, JetBrains AI Assistant, and Antigravity.

What sets Chrome DevTools MCP apart from plain browser automation

Many browser automation tools limit themselves to clicks, input, and reading page content. Per the documentation, Chrome DevTools MCP goes further and brings genuine debugging and performance capabilities normally only accessible through the Chrome DevTools UI itself. A coding agent can therefore not just operate a page, but also measure why it loads slowly, which network requests fail, or which errors show up in the console — information that's often more decisive for debugging web applications than pure interaction ability.

Typical use cases

The server particularly suits tasks where an agent needs to understand the actual state of a web application in the browser: automatically catching performance regressions after a code update, tracing a layout bug through console messages, checking whether a page remains usable under network throttling, or compiling reproducible bug reports with trace data for a development team. For plain content retrieval without interaction or performance questions, a simpler fetch server is often the lighter alternative.

Mind the privacy defaults

Important for production use: the server collects anonymous usage statistics for Google by default and, with performance tools enabled, can send trace URLs to the Chrome User Experience Report API. Both behaviors can be disabled via dedicated flags. Anyone running the server in an enterprise setting or against sensitive internal applications should deliberately review and, where appropriate, disable these settings before production use.

An LTS version of Node.js, npm, and a current stable version of Chrome (or Chrome for Testing). An MCP client with stdio support.

Start the server directly via npx and add the command to the client's MCP configuration. A global installation is not required.

For simpler tasks with a smaller tool set, add --slim (optionally combined with --headless). Further options such as --browserUrl, --isolated, --channel, --executablePath, --viewport, or --proxy-server are documented in the official Configuration Guide.

npx -y chrome-devtools-mcp@latest

No login required. The server launches a local Chrome process and communicates with the client over stdio.

Requires permission to launch and control a Chrome process, plus network access to the visited websites. The client gains access to page content, console output, network data, and performance traces of the controlled browser.

The server exposes browser content, console messages, network data, and performance traces to the MCP client; per the documentation, sensitive or personal information that should not be shared with the client should be avoided. Performance tools can send trace URLs to the Google CrUX API (disable via --no-performance-crux). By default, Google also collects anonymous usage statistics to improve the tool, which can be disabled via --no-usage-statistics or the CHROME_DEVTOOLS_MCP_NO_USAGE_STATISTICS environment variable.

An agent controls a real browser and, with an authenticated session, can trigger real actions on websites or read confidential content. Page content flows into the model's context and should be treated like any untrusted external text. Run it only with the permissions actually needed and, where possible, in an isolated browser instance.