跪拜 Guibai
← All articles
Frontend · Artificial Intelligence · Testing

Google's chrome-devtools-mcp Hands AI Agents the Chrome DevTools They Were Missing

By 狂师 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

AI agents can now handle frontend debugging end-to-end, from opening a browser to reading its internals and proposing fixes, so the copy-paste iteration cycle disappears. But full browser exposure and default telemetry make it a tool to configure carefully before pointing it at anything sensitive.

Summary

Google's ChromeDevTools team released chrome-devtools-mcp in September 2025, and it has already passed 56,000 stars. The MCP server wraps Chrome DevTools' debugging stack into a single npm package, letting AI agents connect to a real Chrome browser and work with the same diagnostic data a human developer sees. Agents open pages, click, simulate mobile or throttled networks, read console errors, inspect network requests, capture performance traces, and run Lighthouse audits, replacing the copy-paste loop of a developer manually exporting errors back to the chat.

The package ships 50+ low-level tools and seven built-in skills that guide agents through specific debugging workflows, from accessibility checks to LCP optimization and memory leak diagnosis. A plugin mode loads all skills alongside the MCP server; a raw JSON config only gets the atomic tools. Setup needs Node.js LTS and a current stable Chrome, and a CLI mode works without an agent. Two caveats: the agent sees everything in the browser, and usage statistics go to Google by default. Both can be managed with flags.

Takeaways
— 56,390 GitHub stars as of 2025-11-15; project open-sourced in September 2025, latest version v1.13.0.
— Built on TypeScript, Puppeteer, and Chrome DevTools Protocol; single npm package with MCP server and CLI.
— Exposes 50+ atomic tools organized into three layers: browser interaction, diagnostics collection, and decision orchestration.
— Ships seven built-in skills: chrome-devtools, chrome-devtools-cli, a11y-debugging, cookie-debugging, debug-optimize-lcp, memory-leak-debugging, and troubleshooting.
— Skills are bundled with the package, not separate installs; plugin mode needed to load them, plain MCP config only has tools.
— Requires Node.js LTS and a current stable Chrome; only Google Chrome and Chrome for Testing are officially supported.
— Flags: --memoryDebugging enables memory snapshot tools, --slim reduces tools, --no-usage-statistics disables telemetry, --no-performance-crux disables CrUX API calls.
— CLI usage: chrome-devtools navigate_page, take_screenshot, and stop commands work without an AI agent.
Conclusions

The most useful agent tools are the ones that give visibility, not just actions; an agent that can read console, network, and trace output produces debugging results that outperform one restricted to clicking.

Google packaging its DevTools analysis engine as MCP tools is a model other teams with specialized debugging stacks can copy, turning internal expertise into reusable agent capabilities.

The default telemetry and unrestricted browser access are easy to overlook during setup, but they change the risk profile for teams considering agent-driven debugging against internal systems.

Concepts & terms
MCP (Model Context Protocol)
An open protocol that lets AI agents call external tools and data sources through a standardized server interface. Here it lets agents interact with Chrome DevTools capabilities.
Chrome DevTools Protocol (CDP)
The protocol Chrome exposes for inspecting and controlling the browser, enabling automation tools like Puppeteer.
Puppeteer
A Node.js library that controls Chrome or Chromium over CDP, used here as the foundation for the MCP server.
Skill vs Tool
Inside chrome-devtools-mcp, tools are atomic operations like opening a page or reading console output; skills are predefined multi-step debugging workflows that show an agent how to combine tools to solve a class of problems.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗