Chrome DevTools MCP with Atomic Agent

chrome-devtools-mcp is the official MCP server from the Chrome DevTools team. It drives a real Chrome browser through Puppeteer and gives an agent DevTools-level access: performance traces, console messages with source-mapped stack traces, network requests, screenshots and page automation. It runs locally over stdio.

Atomic Agent already has built-in browser tools for browsing and filling pages. Add this server when you need DevTools itself: profiling a slow page, reading console errors, inspecting failed requests.

Config

{
  "name": "chrome-devtools",
  "description": "Chrome DevTools: traces, console, network",
  "transport": {
    "kind": "stdio",
    "command": "npx",
    "args": ["-y", "chrome-devtools-mcp@latest"]
  }
}

You need Node.js (LTS) with npm, and current stable Chrome. No account or token is needed. The server starts Chrome on the first tool call that needs a browser, not when it connects.

Optional flags, added to args after the package name (from the project README and its configuration guide):

Flag	What it does
--no-usage-statistics	Turns off the usage statistics Google collects by default for this server
--no-performance-crux	Stops performance tools from sending trace URLs to the Chrome UX Report API
--slim	Exposes a reduced tool set for basic browser tasks, which keeps the prompt shorter on small local models
--headless, --isolated	Run Chrome without a window, or with a temporary profile

For example: "args": ["-y", "chrome-devtools-mcp@latest", "--no-usage-statistics", "--isolated"].

Where to add it:

• Desktop app: Settings › Connections › MCP servers › Add server, paste the object, press ⌘Enter (Ctrl+Enter on Windows and Linux), then Restart on the new row.
• Terminal UI: /mcp add, paste the object. It connects right away.
• By hand: add the object to mcp.servers in ~/.atomic-agent/config.json and restart the agent.

The paste box also accepts the snippet from the project README as is ({"mcpServers": {"chrome-devtools": {"command": "npx", ...}}}).

What to ask the agent

Tools arrive as mcp.chrome-devtools.<tool>. Some prompts that work well:

• “Check the performance of http://localhost:4321 and list the three biggest problems with LCP, with the elements involved.”
• “Open http://localhost:3000/checkout, click Pay, and tell me every console error and failed network request that follows.”
• “Take a screenshot of our pricing page at mobile width and tell me what overflows.”
• “Our signup form returns 500 locally. Reproduce it in Chrome, read the failed request and response, then find the handler in this repo.”

The last prompt is where a local agent helps most: the browser shows the failure, and the agent’s own file and search tools find the code behind it.

Privacy and approvals

• The browser is exposed to the agent. The project README warns that the server lets MCP clients inspect, debug and modify anything in the browser instance. Use --isolated or a separate profile, and do not run it against a browser signed in to accounts you would not hand to the agent.
• Data sent to Google. By default the server sends usage statistics to Google, and performance tools may send trace URLs to the CrUX API. Turn these off with the flags above. This is separate from Atomic Agent’s own analytics.
• What asks for approval. Tools the server does not mark read-only ask before they run, so navigation and clicks wait for your answer in the TUI or the desktop approval card. Do not set trust to pure_read for this server. At approval level 5 (full trust) nothing asks.
• No tokens involved, so nothing secret is stored in config.json for this one.

Troubleshooting

spawn npx ENOENT in the desktop app. The desktop app on macOS (launched from Finder or the Dock) does not inherit your shell’s PATH. Use the full path from which npx as command.

connect timeout after 15000ms the first time. npx -y chrome-devtools-mcp@latest downloads the package on first run. Run that command once in a terminal, stop it, then press Restart.

Chrome does not open. It opens on the first browser tool call, not on connect. If it still fails, check that stable Chrome is installed; the project officially supports Google Chrome and Chrome for Testing.

Trace results look cut off. Tool results reach the model capped at about 8,000 characters. Ask for the top insights rather than the full trace.

Calls time out on heavy pages. Each tool call has 60 seconds in Atomic Agent. Profile a single page load rather than a long session.

FAQ

Is this the official Chrome DevTools MCP?

Yes. It is published by the Chrome DevTools team as chrome-devtools-mcp on npm and GitHub, under Apache-2.0.

Do I need it if Atomic Agent already has a browser?

Not for browsing, reading pages or filling forms. You need it for DevTools features: performance traces, console and network inspection.

Does it work on Windows?

The server is a Node package. Atomic Agent routes a bare npx command through cmd.exe on Windows, so the same config works.

---

Checked on 22 Oct 2026 against Atomic Agent v0.6.10 and Desktop v0.0.4, and the chrome-devtools-mcp README.