In 2025’s developer landscape, AI-assisted coding has evolved from an experimental feature into a fundamental part of the toolchain.
Among the most intriguing ecosystems, the combination of OpenAI Codex CLI and Chrome DevTools MCP (Model Control Protocol) is redefining how we collaborate with AI during software development.

But let’s be honest — every futuristic tool eventually hits that one frustrating error message:

“MCP client for chrome-devtools failed to start: program not found.”

If you’ve seen this line flash across your terminal, you’re in good company.
In this article, we’ll dive into what’s really happening under the hood, how to fix it, and how to unlock the true potential of Codex + DevTools integration.

🧩 1. Scene Setup: The Codex CLI Boot Sequence

When you launch Codex CLI using:

codex

or list available MCP clients:

codex mcp list

you might see something like this:

──────────────────────────────────────╮
│ >_ OpenAI Codex (v0.48.0-alpha.3)       │
│                                         │
│ model:     gpt-5   /model to change     │
│ directory: F:\codex-cli\finohopetex.com │
╰─────────────────────────────────────────╯

■ MCP client for `chrome-devtools` failed to start: program not found

This message means Codex CLI tried to initialize an external MCP client,
but couldn’t find the corresponding executable program on your system.

🔍 2. Understanding MCP: The Bridge Between Codex and the Real World

MCP (Model Control Protocol) is the communication layer between Codex CLI and external “execution environments” like Chrome, VSCode, or Docker.

Think of it as Codex’s way to act — not just talk.
With MCP, Codex can control browsers, inspect pages, run scripts, or even automate full debugging workflows.

Here’s what Codex typically tries to register at startup:

So when you see:

MCP client for chrome-devtools failed to start: program not found

…it simply means the CLI couldn’t locate the Chrome DevTools bridge it needs.

⚙️ 3. The Root Cause: Missing Chrome Path or Misconfigured Environment

Most of the time, this issue boils down to one simple problem —
Codex CLI doesn’t know where Chrome is installed.

By default, the DevTools MCP uses Puppeteer under the hood to launch Chrome,
which requires an environment variable to locate the Chrome binary:

PUPPETEER_EXECUTABLE_PATH="C:\Program Files\Google\Chrome\Application\chrome.exe"

You can verify it by running:

echo %PUPPETEER_EXECUTABLE_PATH%

If it returns nothing, the environment variable isn’t set correctly.

🧰 4. Step-by-Step Fix: Making Chrome DevTools MCP Work

✅ Step 1. Verify Your Chrome Installation Path

On Windows, the default Chrome path is usually:

C:\Program Files\Google\Chrome\Application\chrome.exe

If it differs, manually set it via:

setx PUPPETEER_EXECUTABLE_PATH "C:\Program Files\Google\Chrome\Application\chrome.exe"

💡 Tip: Use setx instead of set to make the variable persist permanently.

✅ Step 2. Install the Chrome DevTools MCP Package

Codex CLI relies on a Node.js bridge module:

npm install -g chrome-devtools-mcp

Or simply run via npx (recommended):

npx -y chrome-devtools-mcp@latest

Once installed, verify Codex recognizes it:

codex mcp list

Expected output:

chrome-devtools   enabled   connected

✅ Step 3. Start Codex in Debug Mode

Launch Codex CLI with debug logs:

codex --debug

You should now see:

[MCP] chrome-devtools connected
[Codex] Ready to receive commands...

At this point, you can execute Codex commands like:

codex run "Open Chrome console and inspect the DOM structure"

Codex will automatically invoke DevTools through Puppeteer and perform the requested actions.

🧠 5. Beyond Debugging: Codex CLI as a Multi-Agent Developer System

Under the hood, the Codex + MCP ecosystem is a multi-agent system that connects:

• 🧩 Codex Core — Natural language reasoning & task orchestration
• ⚙️ MCP Clients — Execution bridges (DevTools, Shell, VSCode, Docker)
• 🖥️ Target Environment — The actual place where commands run

This architecture enables some surprisingly powerful capabilities:

• Automatically scrape and analyze live web pages
• Generate and run site performance audits
• Execute full browser automation via AI instructions
• Debug front-end interactions with natural language commands

In short, Codex CLI isn’t just a command-line interface — it’s an AI-powered development layer.

💬 6. FAQ: Common Issues and Workarounds

Q1: Why do I see “Unsupported Auth” errors?
A: Some MCP clients (like Chrome DevTools) don’t require authentication.
You can safely ignore this or disable Auth checking in .codex/config.

Q2: Can I use this setup on macOS or Linux?
A: Absolutely. Just update the Chrome path accordingly, for example:

export PUPPETEER_EXECUTABLE_PATH="/Applications/Google Chrome.app/Contents/MacOS/Google Chrome"

Q3: Codex still can’t find the MCP package. What should I do?
A: Make sure the chrome-devtools-mcp module is globally installed or accessible within your project directory.

🧭 7. Final Thoughts: Debugging Is the Doorway to Mastery

The combination of Codex CLI + Chrome DevTools MCP marks the beginning of a new AI-driven developer era —
one where debugging isn’t a chore, but a collaboration between human intent and machine execution.

That “failed to start” message?
It’s not just an error — it’s an invitation to understand how your tools think.

The next-generation developer isn’t someone who writes code —
but someone who codes with AI.

📘 Key References

• OpenAI Codex CLI Documentation (GitHub)
• Chrome DevTools Protocol Guide (Google)
• Puppeteer API Reference
• Node.js Environment Variables