Chrome DevTools MCP

Chrome DevTools MCP is the rare case of a vendor building the bridge to its own product and then open-sourcing it: Google's DevTools team ships this as an Apache-2.0 repo with roughly 54k stars and 1.5k commits, which tells you it's an actively maintained project rather than a demo. The design bet is that a coding agent debugging a web app needs the same surface a human engineer uses — console, network, screenshots, performance traces — not a DOM snapshot alone. Three things make that bet credible. First, console messages arrive with source-mapped stack traces, so the agent reasons about your source rather than bundled output. Second, automation runs on Puppeteer and automatically waits for action results, which removes the classic failure mode of an agent clicking before the page settles. Third, performance work isn't limited to lab numbers: trace recordings can be paired with Google's CrUX API for real-user field data, and --no-performance-crux turns that lookup off. The operational surface is small and honest. Install is one npx line in your MCP client; --slim trims the toolset to basic navigation and interaction; --headless and --isolated cover CI and throwaway profiles; the browser starts automatically when a tool needs one. The repo also carries plugin folders for Claude, Cursor, Gemini and Kimi tooling, so setup inside those clients is more than a copy-paste of JSON. Where it bites: the server exposes the browser instance's contents to your MCP client, so a logged-in session on a sensitive account is readable by that client — treat it as a hard boundary, not a footnote. Telemetry is on by default (tool invocation success rates, latency, environment info) and needs --no-usage-statistics or CHROME_DEVTOOLS_MCP_NO_USAGE_STATISTICS to stop. Only Google Chrome and Chrome for Testing are officially supported; other Chromium builds may work but aren't guaranteed. And it is not a high-concurrency scraping tool — it drives a real browser, one session at a time. Who should care: front-end and full-stack engineers already running an MCP client who want the agent to debug against live DevTools rather than a text dump. If you need Firefox and WebKit coverage, Playwright MCP covers more browsers but goes less deep on DevTools-grade profiling. If you just need scripted clicks, a lighter automation server does the job.