COLT converts natural language instructions into browser automation. You say "Create a user with email admin@test.com and admin role" — it executes it on a live browser and exports a standalone Playwright script. How it works:

1.Discover — crawls your web app autonomously, maps every page, form, modal, and element 2.Index — LLM-summarizes each state into a vector search index (discover once, run unlimited tasks) 3.Execute — ReAct agent loop drives the browser in real time with self-healing on failures 4.Record — exports reusable Playwright scripts, pytest tests, and typed Python functions

It handles Shadow DOM, cross-origin iframes, infinite scroll, pagination traps, and modals — the stuff that breaks most automation tools. What it's not: Not a consumer AI browser. Not an RPA tool. It's for engineering teams who want to generate E2E tests or automate internal tool workflows without writing selectors by hand. The key difference from other AI browser agents: they start from scratch every execution. COLT builds a persistent knowledge base of your app, so task #100 is as fast as task #1. Built with Python, Playwright, and ChromaDB. Works with Groq, OpenAI, Anthropic, or local Ollama models. Currently in closed beta — launching soon. Would love feedback on the approach, especially from anyone doing browser automation at scale