I wrote Tabrelay, a small MIT-licensed tool that lets scripts and AI agents drive the Chrome you already use and are logged in to: https://github.com/aaronsuns/tabrelay
Why I made it: I didn’t want a second “automation browser” to log in to again (2FA, new-device checks, sites that treat it like a bot), and I didn’t want a debugging port on my everyday browser that every local program can talk to.
How it works: a Chrome extension plus one Python file (standard library only, no npm, no build step). The command sends jobs (open a page, run JavaScript, click, type, upload a file, screenshot) over 127.0.0.1 to the extension, which runs them with the Chrome DevTools Protocol. Jobs run in background windows so it doesn’t steal focus. The extension only touches sites on an allow-list that can only be edited on its own options page, and it only obeys jobs carrying a local token. It’s a CLI, not an MCP server, so any agent that can run a shell command can use it.
Honest limits: it’s a week old, Chrome/Chromium only, loaded as an unpacked extension (not on the Web Store), a page can tell a debugger is attached, and it has no auto-waiting locators like Playwright. For tests and servers, use Playwright or Puppeteer.
AI disclosure: most of the code, tests and docs were written with Claude Code. I designed it, steer it, review what goes in, and use it myself every day.
Guides and comparisons: https://tabrelay.zexa.se/
What a catostrophically bad idea. Agents are sandboxed for good reason, delusional slop slinger.
Absolutely not, ever.
No.
Ew



