Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseCode / SDK · ActionableGitHub Trending AI→ original

Google gives Claude and Cursor a DevTools panel: the agent finds out for itself why a site is slow

An AI agent edits a website blind: it can't see what failed in the console or why a page loads slowly. The Chrome team has released chrome-devtools-mcp, a server that connects Claude, Cursor and Copilot to a real Chrome with DevTools tools (the same panel you open with F12). The agent records a load trace, checks requests and the console, and takes screenshots. That means the "write, test, fix" loop can be handed to the agent in full.

AI-processed from GitHub Trending AI; edited by Hamidun News
Google gives Claude and Cursor a DevTools panel: the agent finds out for itself why a site is slow
Source: GitHub Trending AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceGitHub Trending AI ↗
Editor-in-Chief
Key metric #1Works as an MCP server and as a CLI utility for the terminal
Setupadd the chrome-devtools-mcp@latest string to the client config
Key metric #3Google collects usage statistics by default; a flag turns this off

Anyone who has asked an AI agent to fix layout or speed up a page knows the typical failure: the agent makes changes, confidently reports "done," and the browser is broken. The reason is simple: it writes code but cannot see the result. The Chrome team has addressed this gap with the ChromeDevTools/chrome-devtools-mcp project, which gained around 60 stars on GitHub Trending within a day. It is a server that gives the agent the same tools you get by pressing F12: DevTools, Chrome's built-in developer panel with a console, network requests and a profiler. Everything connects through MCP (Model Context Protocol), a standard "socket" through which AI clients such as Claude, Cursor or Copilot call external tools.

What it does

Chrome DevTools for agents runs as an MCP server. You can connect Claude, Cursor, Copilot, Antigravity and other clients to it. The agent gains the ability to control a live Chrome browser and look inside it.

According to the project description, the feature set includes:

  • recording a performance trace (a detailed log of page loading) and analyzing it for practical findings: what exactly is slowing things down;
  • analysis of network requests;
  • page screenshots;
  • reading console messages along with stack traces mapped to the source code, not to the minified bundle;
  • automating browser actions with a wait for results, so the agent doesn't click a button that hasn't rendered yet.

A simple example from the documentation: tell the agent "Check the performance of https://developers.chrome.com." The client will open the browser and record a trace. This also serves as a quick check that the installation worked. The browser launches only when the agent calls the tool that needs it; connecting to the server does not open Chrome by itself.

Installation, CLI and skills

Setup follows the standard MCP pattern: you add the chrome-devtools-mcp@latest package to the client config. The @latest suffix means the client always pulls the newest version. For Claude Code and Gemini CLI there are installation options via a plugin or extension, and instructions for Cursor and VS Code are in a separate guide to client configurations.

The package also includes a console utility, chrome-devtools. It is for cases where MCP is not needed: you can control the browser from the terminal and write shell scripts. Commands look like navigate_page with an address or take_screenshot with a file path. It is installed globally via npm.

The authors have also added "skills": hints that teach the agent how to use DevTools. They cover:

  • general debugging and page automation;
  • working through the CLI;
  • auditing and fixing accessibility issues;
  • debugging cookies, sessions and authentication;
  • improving Largest Contentful Paint, that is, how quickly the main block of a page appears;
  • finding memory leaks;
  • fixing connection problems between MCP and Chrome.

An important caveat: skills describe the tools of a specific version. After updating the server or CLI, the skills need to be updated too, or the agent may call parameters that have already changed. Plugin managers update them automatically, and the Gemini CLI extension can be installed with auto-update. If skills were copied by hand, they must be copied again.

For complex scenarios there are additional settings: parallel sessions, persistent user data directories, a non-standard browser path, connecting to an already running Chrome, and debugging on Android. For simple tasks there is a flag that limits the toolset. For developers building their own agent products, the authors recommend building an embedded browser subagent on top of Chrome DevTools for agents. The reference implementation is the browser agent in Gemini CLI.

Risks and privacy

Be careful here. The server exposes browser contents to MCP clients, and they can view, debug and modify any data in the browser or DevTools. The authors explicitly ask users not to share what they don't want the agent to see. In practice, this means: don't debug through an agent using a profile with working sessions, bank logins and personal email; use a separate clean profile instead.

There is also data collection. Google collects usage statistics by default: the frequency of successful tool calls, latencies and environment information. You can turn this off with a flag when starting the server or with the CHROME_DEVTOOLS_MCP_NO_USAGE_STATISTICS environment variable. Opting out of Chrome's own metrics does not automatically disable this tool's metrics, and vice versa. In addition, performance tools may send addresses from the trace to the Google CrUX API to show real-user field data alongside lab data. The server also periodically checks npm for updates; this can be disabled with the CHROME_DEVTOOLS_MCP_NO_UPDATE_CHECKS variable.

Only Google Chrome is officially supported. Other Chromium-based browsers may work, but there are no guarantees.

What it means in practice

For a front-end developer, this is a step from "the agent wrote code" to "the agent wrote it, tested it and saw the error itself." A couple of ready-made scenarios: have the agent find the cause of a slow LCP, check a failed network request, or look for a memory leak. For founders building agent products, the project offers a ready-made foundation for a browser subagent. Test everything on a throwaway profile first, and decide right away whether you want telemetry.

Story

Key Questions Answered

What it does?

Chrome DevTools for agents runs as an MCP server. You can connect Claude, Cursor, Copilot, Antigravity and other clients to it. The agent gains the ability to control a live Chrome browser and look inside it. According to the project description, the feature set includes: recording a performance trace (a detailed log of page loading) and analyzing it for practical findings: what exactly is slowing things down; analysis of network requests; page screenshots; reading console messages along with…

What are the key details on Installation, CLI and skills?

Setup follows the standard MCP pattern: you add the chrome-devtools-mcp@latest package to the client config. The @latest suffix means the client always pulls the newest version. For Claude Code and Gemini CLI there are installation options via a plugin or extension, and instructions for Cursor and VS Code are in a separate guide to client configurations. The package also includes a console utility, chrome-devtools.

What are the key details on Risks and privacy?

Be careful here. The server exposes browser contents to MCP clients, and they can view, debug and modify any data in the browser or DevTools. The authors explicitly ask users not to share what they don't want the agent to see. In practice, this means: don't debug through an agent using a profile with working sessions, bank logins and personal email; use a separate clean profile instead. There is also data collection.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…