bettyguo/browser-skills

Browser-using AI agents are good at reasoning but slow at routine web interactions. They re-figure out how to dismiss a cookie banner every session, get stuck on infinite scroll feeds, and burn extra resources on calendar widgets. Browser-skills is a bundle of 15 reusable recipes, called skills, that handle these common repetitive patterns so agents do not have to solve them again each time. Each skill is a SKILL.md file containing step by step instructions for a specific web interaction. The skills are selector driven by default, with vision used only as a fallback, so most complete in a few hundred milliseconds and use no extra model tokens. The 15 skills cover patterns such as dismissing cookie banners and newsletter popups, handling modal dialogs and exit intent overlays, verifying a page has finished loading, detecting captchas without solving them, filling multi step forms, uploading and downloading files, login flows, extracting paginated tables, scrolling infinite feeds, searching and filtering, date picker widgets, and searchable dropdowns. You install the package via pip and run a single command to wire it into your AI tool of choice, with the README naming Claude Code, Cursor, Codex CLI, and Continue as examples. There is also an MCP server that exposes the skill library over a standard protocol, letting agents start a browser session, list available skills, check which ones apply to the current page, invoke a skill by name, and take a screenshot. The project is written in Python and uses Playwright for browser control. It explicitly does not solve captchas, evade anti-bot systems, or store login credentials, and it respects robots.txt with a human paced request rate. You can author new skills yourself by writing a SKILL.md file in the documented format. The license is MIT.