Google's Chrome DevTools MCP Lets AI Agents Debug in Chrome

Share

Google's Chrome DevTools MCP 1.13.0 adds logpoints that read variables in live Chrome without editing source, and network mocks that serve local files or test-server data for API responses. Mocks are off by default, and aborting requests comes in the next version, plus a prompt to hand to your AI agent.

Machine translation. Read the original

SourceChrome DevTools MCP(GitHub Release)Reported Nov 11, 2026Not yet tested

Google's official Chrome DevTools MCP released version 1.13.0 on November 11. This update lets AI agents set logpoints in real Chrome, so they can see runtime variables without editing the source code. It can also swap the response from a given API for a local file or for data from a test server, so you can test different cases without touching frontend code.

What it is

Chrome DevTools MCP is an open-source project maintained by the Chrome team (Apache-2.0, about 57,000 GitHub stars). It uses MCP, a general interface that lets AI call external tools, to hand Chrome's developer tools to coding agents such as Claude Code, Codex, Cursor, and Copilot. Once connected, an agent can open web pages, click buttons, take screenshots, read console errors, inspect network requests, and record performance data. It works much like pressing F12 yourself.

Before this, when an agent debugged a frontend bug, it usually stuffed a pile of console.log calls into the code, refreshed the page, read the output, and then had to remember to delete them. This update adds two new pieces.

What's new

Logpoints. The agent tells Chrome, "At line 42 of src/app.ts, print user and cart.total." When the code reaches that line, the values are printed to the console. The page doesn't pause, and not a single line of source changes. Bundled projects can use original file paths as long as a source map is available (the file a bundler generates to map minified code back to the original source). For now this only works for scripts in the page's main frame. Code running in Web Workers or cross-origin iframes can't get logpoints yet.

Network mocks. The agent can pick requests to a given URL and have the page receive the contents of a local file, or the data returned by another server such as a test environment. The replacement still applies after the page reloads, and if the local file changes, it is read again. This suits cases like "what does the page look like when the API returns an empty list" or "does the flow still work with test data."

This feature is off by default. You have to add --categoryNetworkMocks to the startup arguments to make it appear, and then restart the MCP server. While replacements are active, the browser cache is disabled, so performance recordings are not accurate. The tool will warn you about this.

Some people read this update as "it can simulate API failures," but strictly speaking one step is still missing. Making a request fail outright (abort) was merged into the code a few hours after 1.13.0 was released, so it will ship in the next version.

What it means for you

  • Less code churn when agents debug: the agent reads variable values straight from logpoints. There's no back-and-forth adding and removing logs, and no stray console.log left in to ship to production.
  • Test edge cases without backend help: if the API isn't finished, or test data is hard to create, swap the response for a JSON file first and have the agent check how the page handles it.
  • The whole frontend debugging toolkit in one tool: console errors, network requests, performance analysis, breakpoints, and response replacement now all live in the same tool.

Getting started

If you already installed it, check that your MCP config says chrome-devtools-mcp@latest, then restart and you're on the new version. If not, hand the prompt below to your AI agent.

Help me connect Google's official Chrome DevTools MCP (project at https://github.com/ChromeDevTools/chrome-devtools-mcp) to the AI agent I'm using now. Do it in this order:

1. First check whether this computer has Node.js (LTS version) and the latest stable Chrome. If either is missing, tell me first and ask before installing anything.
2. Following the project README's setup instructions for this agent, add chrome-devtools-mcp@latest, and add --no-usage-statistics to the startup arguments (this turns off usage statistics, which are on by default).
3. Ask whether I want to turn on network response replacement. If I do, add --categoryNetworkMocks.
4. After restarting, test with a public web page (for example https://developers.chrome.com): open the page, take a screenshot, and list the console messages to confirm it works.

From then on, when I ask you to debug my local project:
- Prefer logpoints to read variables. Don't add console.log to the source code.
- Before replacing an API response, tell me which URL you'll replace and what you'll replace it with. After debugging, delete all the replacements.
- Don't open sensitive sites I'm logged into, such as online banking or email.

After each step, tell me the result in plain language, then stop and wait for my confirmation.

Making "no console.log" and "delete the replacements after debugging" into firm rules matters because these are the two things most likely to leave a trail. One ships debug code in a commit. The other leaves you looking at fake page data without realizing it.

Notes

  • This tool gives the AI agent access to what's in your browser so it can read and change things. The official README also warns not to process personal information you don't want shared. Debug in a separate browser window with a test account.
  • Google collects usage statistics by default (call success rates, timings, and environment info). To opt out, add --no-usage-statistics. Performance analysis also sends the page URL to Google's CrUX service by default to fetch real-user data. You can turn that off with --no-performance-crux.
  • Google officially supports only Google Chrome and Chrome for Testing. Whether other Chromium-based browsers work is hit or miss.
  • With network replacements active, performance data is inaccurate. Remove the replacements before recording a performance profile.

Want to learn what MCP is first? See the MCP handbook. Want to give an agent its own browser for side jobs? See this tutorial.

Done with this one? Mark it to track your path.Sign in
Save it for laterSign in
Questions about this piece?Ask AI about this

Comments (0)

No comments yet. Start the conversation.

Sign in to join the discussion.Sign in