Google Chrome with direct AI Agents interface

Chrome 146 has added a way to let AI Agents directly interface with it. The advantage over existing browser automation projects like the Playwright MCP is that this lets you control already running, potentially logged-in, browser sessions.

The feature flag to be enabled for this is “Allow remote debugging for this browser instance”:

chrome://inspect/#remote-debugging

The official way is to use it through the Google DevTools MCP server but that’s not a requirement. In fact, Codex can be prompted as simple as:

Return three URLs I have currently open in Chrome. (Note: This Chrome uses DevToolsActivePort instead of CDP discovery. Read DevToolsActivePort from the profile and connect to ws://127.0.0.1:<port><path>)

This will show a warning a warning in Chrome about the browser being accessed externally and will only proceed when confirmed. If Codex accesses Chrome several times, the same warning will be shown each time. A workaround to this is having Codex run a long-lived background process that holds the connection rather than dropping it after each action. Petr Baudis has created a SKILL for that, but that is built on Node.JS. The same can be achieved without Node however, ad-hoc done by Codex itself:

Return three URLs I have currently open in Chrome. (Note: This Chrome uses DevToolsActivePort instead of CDP discovery. Read DevToolsActivePort from the profile and connect to ws://127.0.0.1:<port><path>. Don’t connect directly - instead, create and run a long-lived generic Chrome DevTools helper process (CDP pass-through; PowerShell or Python venv) that will keep the TCP connection open for reuse.)

This is slow for the first request when Codex has to create the helper process first, but will be re-used on subsequent requests. Caution: this helper may not have own guards - like access controls - in place, so only use this in a trusted environment.