Richárd Hruby builds browser-agent software at Anon. His Login Machine project and accompanying essay show a login workflow in which a model interprets the screen while credentials travel directly to the browser.[3]
Start with Login Machine →
X Contributions
Separate screen interpretation from secret input
Login Machine extracts form-relevant HTML and captures a screenshot, then asks a model to classify the current screen and return structured fields and locators. The user supplies requested input, which Playwright sends directly to the browser DOM. The model describes the form rather than receiving the entered credentials. After each action, the loop observes the new page instead of assuming a fixed sequence.[4]
Check proposed actions against the live page
Hruby’s essay explains validating model-generated locators before acting. Missing locators become error context for another attempt, while distinct schemas represent credential forms, choices, magic links, waiting, blocking overlays and successful login. The demo excludes CAPTCHA handling. Hruby credits the project to himself and Jesse at Anon, so the selected work preserves collaboration rather than assigning sole authorship.[3]
Tools, writing & talks
Tools
Start here
Login Machine
github.comAn open-source browser-login loop with structured screen classification and credential isolation.
Read the architecture and credential-isolation sections before trying the demo. Local use needs Node.js and configured Anthropic and Browserbase credentials; the accompanying essay identifies CAPTCHA handling as outside the demo.
Cocreator with Jesse at Anon[4]Original source
Writing, research & talks
- I replaced 100 login scripts with a browser agent loophrubyonrails.substack.com
A first-person implementation account of screen schemas, locator validation and retry context.
Author[3]