Drive a browser

Most browser automation tools require a headless browser, separate logins, and a lot of code. But starting with Chrome 144, your AI agent can drive your real browser — the one you use every day, with all your tabs open and all your accounts logged in.

This works through the Chrome DevTools MCP server, which connects OpenCode to Chrome’s remote debugging protocol. Chrome shows a permission dialog each time and displays a banner while the agent is connected, so nothing happens silently.

Set up Chrome DevTools MCP

Two things need to happen:

Enable remote debugging in Chrome — navigate to this URL in Chrome and enable it:

chrome://inspect/#remote-debugging

Add the MCP server to your OpenCode config — add this to your global config (~/.config/opencode/opencode.jsonc):

{
  "mcp": {
    "chrome-devtools": {
      "type": "local",
      "command": ["npx", "-y", "chrome-devtools-mcp@latest", "--autoConnect"]
    }
  }
}

Restart OpenCode after making this change.

Start in Plan Mode

Before letting the agent interact with your browser, switch to Plan Mode. In this mode, the agent describes what it intends to do and waits for your approval before acting. This is a good safety practice — your browser has access to your accounts, your data, and your sessions. You want to review actions before they happen.

The killer feature: your existing session

The reason this is powerful is that your agent inherits your existing Chrome session. You’re already logged in everywhere. That opens up sites that were never designed to be automated — sites with no API, no CLI, and a complex authenticated UI:

• Check your bank transactions and categorize spending
• Summarize a paywalled article you subscribe to
• Extract assignment due dates from a school portal
• Run a Lighthouse accessibility audit on any page
• Fill out a long form with your agent walking you through it
• Compare product details across multiple tabs

What you’ll do

You’ll start simple — take a screenshot of your current tab to verify the connection works. Then try a Lighthouse audit. After that, the exercise is open-ended: pick something you actually want to automate and do it.