Hey everyone, I built Sarathi AI Agent after experimenting with different approaches to browser agents. Most current AI browser tools rely on screenshots + vision models. I wanted to see if structured DOM reasoning could make it faster, more deterministic, and easier to debug. So I built a Chrome extension that: Injects unique IDs into visible elements Creates a structured DOM snapshot Lets the LLM return deterministic actions (click, type, navigate, hover, wait) Executes in a loop until completion It currently supports: – Gmail reply drafting – Intelligent form filling – E-commerce flows – Hover-based interactions Still experimental, and I expect edge cases — especially around auth redirects and anti-bot systems. I’d genuinely appreciate feedback from the community: • Where do you see this being useful? • What risks or guardrails should I consider? • DOM-based vs screenshot-based — which do you prefer and why? Thanks for checking it out.

Sarathi AI Agent

Open-Source DOM-Based AI Browser Agent

Open-Source DOM-Based AI Browser Agent