Chrome DevTools MCP — Chrome DevTools in your agent’s hands
It gives a coding agent a live Chrome window. The agent can open pages, fill forms, click, and type.

What it does
It gives a coding agent a live Chrome window. The agent can open pages, fill forms, click, and type.
What makes it useful is simpler: it sees the same DevTools data a human sees.
That means performance traces with Core Web Vitals, network requests, console messages with stack traces, screenshots, heap snapshots, and Lighthouse audits.
The tool reference counts 59 tools in 11 categories. There is also a CLI if you want to run it without MCP.
What it does not do
It only sees Chrome DevTools, so it is not a cross-browser tool. It runs with Chrome DevTools. If you need control over Firefox, Safari, or real mobile devices with another runtime, it will not cover that. It is not a replacement for device farms or for measurements from real users in the field. The data you get is what Chrome can provide, not an all-powerful simulation of every environment.
It does not apply automatic fixes. It sends metrics, traces, heap snapshots, and console logs. The agent can suggest fixes, but it does not apply safe patches, manage deployments, or replace human judgment for UX and accessibility. Do not ask it to “fix it and put it in production” without manual review.
It is heavy and session-centric. It needs Chrome and enough CPU and RAM. The default surface consumes context tokens, and the slim flags reduce the number of tools at the cost of information. It also needs extra setup for authenticated flows and to avoid telemetry to Google. If you want full local, lightweight, or cross-browser testing, you will need a different tool or extra infrastructure.
Who it is for
For the moment when “it works in the terminal” is not enough. When the issue is rendering, network, or speed.
Why it is worth it
First: it is official. The Chrome DevTools team ships it. They announced it on their blog on 23 September 2025. You are not waiting on a third party to maintain the tool.
Second: the way it sees the page. It does not send screenshots to a model to guess. It sends structured data from the DevTools protocol. Metrics, not pixels.
Third: the cost. The default surface burns tokens. That is why slim exists, leaving only 3 tools. And it sends telemetry to Google, with two separate opt-outs.
Installation
Put this in the client config:
{
"mcpServers": {
"chrome-devtools": {
"command": "npx",
"args": ["-y", "chrome-devtools-mcp@latest"]
}
}
}The @latest keeps it on the latest release.
If you want lighter sessions:
"args": ["-y", "chrome-devtools-mcp@latest", "--slim", "--headless"]Only 3 tools, with no window.
Use in each LLM
Claude Code
From the terminal:
claude mcp add chrome-devtools --scope user npx chrome-devtools-mcp@latestThat adds the server for all your projects.
Or as a plugin, which bundles MCP and skills together:
/plugin marketplace add ChromeDevTools/chrome-devtools-mcp/plugin install chrome-devtools-mcp@chrome-devtools-pluginsThen restart Claude Code. Check with /skills.
Codex CLI
codex mcp add chrome-devtools -- npx chrome-devtools-mcp@latestThat adds the server to Codex.
On Windows 11 it needs a different setup. npx does not find Chrome correctly. Set command = “cmd” and raise startup_timeout_ms to 20,000.
Cursor
Settings → MCP → New MCP Server. There you paste the JSON config from above. There is also an install button in the repo.
Where it wins, where it loses
It wins on verification. The agent sees the page with its own eyes, not with imagination.
It loses on weight. It needs Chrome. The default surface eats context. The size flags are not a luxury, they are a need. And telemetry needs care: two opt-outs, not one.
Test
I ran:
npx -y chrome-devtools-mcp@latest --versionIt returned 1.10.1, exit code 0, on Node 26.
npx -y chrome-devtools-mcp@latest --helpIt printed all parameters. There I saw the defaults: usageStatistics true, performanceCrux true, and the –slim and –headless flags available.
What I did not do: connect the server to a client, or open a page. Everything I write about the flow inside the client comes from the documentation.
Verdict
Worth it, if you put it in the right place. It is the clean way to let an agent see and check a page. It is not for every edit. Keep it narrow: slim mode and the right flags.
Score: 8/10.
Also see
In the same category we also tested:
- ty: Astral’s type checker that catches the mistake before it runs
- cf: Cloudflare’s new CLI that does not guess flags
- ast-grep: grep that understands code, not text
Sources:
https://github.com/ChromeDevTools/chrome-devtools-mcp
https://developer.chrome.com/blog/chrome-devtools-mcp
https://github.com/ChromeDevTools/chrome-devtools-mcp/blob/main/docs/tool-reference.md
https://github.com/ChromeDevTools/chrome-devtools-mcp/blob/main/docs/client-configurations.md
https://www.npmjs.com/package/chrome-devtools-mcp
https://api.github.com/repos/ChromeDevTools/chrome-devtools-mcp
Εγκατάσταση
2 ΒΗΜΑΤΑ- npx -y chrome-devtools-mcp@latest --version
- npx -y chrome-devtools-mcp@latest --help