The Chrome team announced the public preview of the DevTools MCP (Model Context Protocol) server . Through MCP, AI-powered coding agents can directly control a real Chrome instance, performing performance tracing , DOM/CSS inspection , console reading, script execution , and real-time debugging on the page. This addresses the long-standing "invisible and intangible" nature of agents in the browser's runtime state.
The official blog and repository provide sample tools (such as performance_start_trace): The agent can launch Chrome, open the target site, record a performance trace, analyze bottlenecks, and provide actionable optimization suggestions. Based on the standardized interface of MCP, DevTools capabilities can be reused by clients such as Gemini, Claude, Cursor, and Copilot , suitable for scenarios such as web performance optimization, UI regression, and end-to-end debugging. Currently in public beta, subsequent features will be continuously improved.
Frequently Asked Questions
Q: What is DevTools MCP?
A: A server exposed via the MCP protocol that allows AI agents to use the debugging and performance capabilities of Chrome DevTools through standard tool calls.
Q: What specific operations can be performed?
A: Start performance tracing, inspect DOM/CSS, read Console, execute JS, record and analyze Trace, and support automated user processes.
Q: What proxies/clients are supported?
A: For clients that support MCP, the official examples indicate that Gemini, Claude, Cursor, Copilot , etc. can all be integrated.
Q: What is the current release phase?
A: Public preview ; the code and npm package are available, and features will be iterated based on feedback.
Q: What is its relationship with the existing CDP (Chrome DevTools Protocol)?
A: MCP is a higher-level unified bridge protocol that standardizes the connection between clients and tools; the underlying layer still utilizes DevTools/Chrome capabilities.
