Building another browser AI
A small project that accidentally became something bigger.
I didn’t plan to build a browser agent, nor did I want to
I just wanted Scrapezy’s data extraction to go a level deeper.
Users started asking for things that only a proper agent could do, complex navigation, interacting with sites, more anti bot detection. So I started looking at the usual suspects: Stagehand, Playwright, browser-use.
They’re all great, but each had limitations. Stagehand and Browser-User depend on CDP (Chromium DevTools Protocol), but I wanted to use Camoufox a Firefox-based anti-detection browser. That means no CDP and suddenly none of the existing tools fit.
So I started building something new.
It borrows from Stagehand and Browser-Use and uses Playwrights ability to use most browsers. The goal is to let you bring any LLM model and any Playwright-compatible browser.
I want something that can cache the LLMs result, and re-use it or generate Playwright script. So you only pay the upfront cost once. It’ll be significantly quicker too. Eventually, I want to blend the caching so it can re-use the cache and handle subtle changes in a page (potentially recaching them) .
The hardest part was getting the agent loop right, that cycle of observe → decide → act → repeat. Prompting still needs work; it’s easy to overload the context or starve it of info. But it’s improving.
When I finally saw it scrape three of the top five Product Hunt launches, I knew it worked.
Right now I’m deciding whether to open-source it or spin it out as a separate product. Maybe it becomes part of Scrapezy. Maybe it’s a standalone agent with a hosted version.
Either way, this reminded me why I build things: to understand how they work.
Just build things. You’ll learn something every time.