Chrome DevTools MCP Server Setup Guide
On this page
When the agent needs DevTools, not just a browser
"The dashboard feels slow" is hard for an agent to act on. With Chrome DevTools MCP connected, it can open the page, record a performance trace, read the Largest Contentful Paint (LCP) breakdown and tell you that a 900 KB hero image is the problem. It's built by the Chrome DevTools team and drives Chrome through the same protocol DevTools uses.
Chrome DevTools MCP or Playwright MCP?
Both open a browser and click things, so it's easy to install the wrong one.
| You want to | Pick |
|---|---|
| Test a user flow: sign up, fill a form, check the result | Playwright MCP |
| Use Firefox or WebKit as well as Chrome | Playwright MCP |
| Find out why a page is slow, with traces and Core Web Vitals | Chrome DevTools MCP |
| Debug failing network calls, console errors, CSS or memory | Chrome DevTools MCP |
| Run a Lighthouse audit | Chrome DevTools MCP |
Plenty of people connect both and let the agent pick. Turn one off if you notice the agent mixing them up.
Add it to your tool
You need a current Node.js LTS and current stable Chrome. Version 1.12.0 was current in November 2026.
Claude Code:
claude mcp add chrome-devtools --scope user -- npx -y chrome-devtools-mcp@1.12.0Cursor, Claude Desktop and other mcpServers clients:
{
"mcpServers": {
"chrome-devtools": {
"command": "npx",
"args": ["-y", "chrome-devtools-mcp@1.12.0"]
}
}
}Codex CLI and Gemini CLI:
codex mcp add chrome-devtools -- npx -y chrome-devtools-mcp@1.12.0
gemini mcp add -s user chrome-devtools npx -y chrome-devtools-mcp@1.12.0On Windows with Codex, start npx through cmd /c and raise startup_timeout_ms, as shown in the Codex guide.
What it can do
Listing version 1.12.0's tools with the MCP Inspector, in headless mode, returned 32 tools. They fall into a few groups:
- Performance:
performance_start_trace,performance_stop_trace,performance_analyze_insight, andlighthouse_audit - Network:
list_network_requests,get_network_request - Debugging:
list_console_messages,evaluate_script,get_css_styles,take_snapshot,take_screenshot,take_heapsnapshot - Navigation and input:
navigate_page,new_page,click,fill_form,type_text,wait_for - Emulation:
emulate,resize_page, to test slow networks or small screens
Try a prompt like: "Open localhost:3000, record a performance trace while it loads, and tell me what's hurting LCP." The agent starts a trace, reloads, stops it and reads the insights.
Useful flags
Add these after the package name in args:
--headlessruns without a window. The default shows Chrome so you can watch.--isolateduses a temporary profile that's deleted when the browser closes.--channelpicksstable(the default),beta,devorcanary.--viewport=1280x720sets the starting window size.--browser-url=http://127.0.0.1:9222attaches to a Chrome you started yourself with remote debugging on, instead of launching a new one.--slimexposes a minimal set of three tools for navigation, scripts and screenshots, if you only need the basics.
Privacy and safety settings
Read these before you use it at work:
- The agent sees the whole browser. Anything open in that Chrome instance, including logged-in sessions, is visible to the agent and to the model provider. Use
--isolated, or a separate profile, for anything sensitive. - Usage statistics are on by default. Turn them off with
--no-usage-statistics, or by setting theCHROME_DEVTOOLS_MCP_NO_USAGE_STATISTICSenvironment variable. - Performance tools may send page URLs to Google's CrUX API to fetch real-user metrics. Turn that off with
--no-performance-cruxfor internal sites.
For an internal app, a sensible setup is:
"args": ["-y", "chrome-devtools-mcp@1.12.0", "--isolated", "--no-usage-statistics", "--no-performance-crux"]Check yourself
Chrome DevTools MCP, using performance_start_trace, performance_stop_trace and performance_analyze_insight.
--isolated so your normal browser sessions aren't exposed, and --no-performance-crux and --no-usage-statistics so internal URLs and usage data aren't sent to Google.
No. Chrome DevTools MCP supports Chrome only. Use Playwright MCP with --browser webkit.