Most AI desktop agents screenshot your screen and send it to a vision model for every action. Clawd Cursor takes a different approach — it uses screen reader accessibility APIs first, falling back to vision only when needed. The result: 80% of tasks need zero LLM calls. It's 6x faster and 30x cheaper than screenshot-based agents. Built with TypeScript, it connects via VNC and uses a smart action router that tries accessibility APIs, then task decomposition, then AI vision as a last resort.
