wtf is reverseloom?

reverseloom is an AI browser agent built to figure out how a website's backend API works, then write a standalone script that talks to that API directly, without needing a browser running at all. The idea is that many websites protect their data with bot detection systems, signed or encrypted request parameters, and other barriers. Instead of a person manually studying obfuscated JavaScript for a day to reverse engineer one of these systems, an AI model drives a real browser, watches the page's network traffic and JavaScript debugger, and works out how a signature or token is generated. It then reproduces that logic in a small JavaScript sandbox and tests it repeatedly to confirm it works from a cold start, before writing a plain Python crawler that reproduces the same requests without a browser attached. The README walks through a real example against a site protected by Akamai Bot Manager: the agent notices a missing header causes a 403 error, sets breakpoints on the relevant JavaScript function, pulls out the two scripts responsible for generating that header, rebuilds the logic in its sandbox, and ships a working crawler that pulled 64 price records and passed five out of five repeated cold-start tests. The tool includes over 30 capabilities grouped into browser automation (clicking, typing, scrolling, dragging), JavaScript reverse engineering through Chrome DevTools Protocol breakpoints and network inspection, a vision tool for locating things like captchas on the page, and a way to hand off control to a human when a login or captcha needs manual input. It works with any Chromium-based browser already installed on the machine (Chrome, Edge, Chromium, or Brave) and does not download its own browser. Setup on Windows is a downloadable executable with its own bundled Python runtime, on macOS, Linux, or for developers, it installs from source with pip and runs as a desktop app or a local web server. Configuration happens through a settings screen or a .env file, where the user picks a model provider (OpenAI, Anthropic, Gemini, DeepSeek, or a local Ollama model) that supports both image input and streaming output. The README warns plainly that one of its tools runs arbitrary shell commands, so it should only be pointed at trusted code paths. The project is licensed under Apache 2.0 and is currently marked alpha status.