Browser automation has always meant starting over—no sessions, no cookies, and websites immediately flagging you as a bot the moment your "browser" boots with an empty profile and headless markers. A new open-source tool called chrome-use flips that script entirely by letting any AI agent tap directly into Chrome instances you're already logged into everything on.
The Architecture: Extension + Native Messaging, Not Debug Ports
The project uses a lightweight Chrome Web Store extension paired with native messaging—a local inter-process channel that bypasses network sockets and tokens entirely. When you run chrome-use open https://x.com/home, it connects to your real browser over this relay, drives tabs via chrome.debugger, and returns results without ever touching the debug port that Chrome 136+ now guards with a blocking consent dialog on every attach. The key differentiator from raw CDP approaches like web-access: native messaging requires a one-time extension install, then zero per-use confirmation. Your agent connects silently in the background while you keep working—new tabs open unfocused in colored per-session tab groups so multiple agents can share one real Chrome without stepping on each other or your own browsing.
Anti-Detection Without JavaScript Patches
When connected to your actual browser, chrome-use injects zero JavaScript modifications. Your fingerprint stays completely genuine because there's nothing to lie about—the stealth comes from native CDP/Chrome overrides rather than re-defined getters that themselves become detectable tells. The tool scores 0% bot on CreepJS (the most thorough fingerprint and lie detector available), passes all checks on bot.incolumitas.com, gets green lights across sannysoft and BrowserScan, and clears Cloudflare Turnstile. Runtime.enable is off by default since a live CDP domain itself becomes a detectable signal—the "runtime leak" that rebrowser and similar tools flag even when attached to your real Chrome.
Site Adapters and Structured Data Without DOM Scraping
Beyond clicking and screenshotting, chrome-use ships a site adapter system that taps into websites' own JSON APIs from inside your logged-in tab—your cookies, same-origin fetch permissions, the site's own modules. It fetches community-built adapters (~145 commands at runtime) for GitHub issues, Reddit search, Bilibili feeds, and more, returning clean structured JSON instead of parsing rendered DOM. The adaptive detector auto-escalates pages guarded by Akamai, PerimeterX, or DataDome to "humanize" mode—curved cursor trajectories with decelerating Bézier paths landing on jittered points inside elements (never dead centre), variable inter-keystroke timing for typing, eased scroll segments. Pages without bot vendors stay instant with zero overhead.
Comparison: Why chrome-use Wins Over Playwright and Claude's Extension
Playwright and Puppeteer boot empty browsers requiring full re-login and flagging immediately on CreepJS. Claude in Chrome only drives Claude itself—not Cursor, Codex, or your own scripts. Raw --remote-debugging-port approaches pop that consent dialog on every connection with Chrome 136+. The extension-native messaging path avoids all three pain points while scoring identical 0% stealth on CreepJS because it genuinely IS your browser.
Installation and Usage
Install via a single shell command: curl -fsSL https://raw.githubusercontent.com/leeguooooo/chrome-use/main/install.sh | sh. Add agent skills with npx skills add leeguooooo/chrome-use for pre-approved bash permissions in Claude Code, Cursor, or Codex contexts. Commands are simple—chrome-use open
The Bottom Line
This is what browser automation should have been from day one—no fresh profiles, no re-authenticating every session, and no debug port consent dialogs interrupting your agent workflows. If you're running any AI-assisted development with browser interactions, chrome-use removes the friction that made Playwright feel like fighting the tool instead of using it.