macOS computer use 1.8x faster, 85% cheaper than cua-driver alone

A new tool called beans-picker has been developed to optimize how AI agents interact with macOS applications. By building on top of the existing cua-driver, the project addresses the high token consumption and latency caused by agents reading full accessibility trees and screenshots for every interaction. Instead, beans-picker extracts a list of candidate actions from the accessibility tree and uses a specialized model, Jev, to select the correct control. Benchmarks indicate that this approach is 1.8x faster and 85% cheaper than using cua-driver alone, with median task times dropping from 49.7 seconds to 28.1 seconds and costs decreasing from $0.388 to $0.060 per run. The tool is designed for background automation, allowing users to continue working while the agent performs tasks. The project is open-source and available on GitHub for those looking to improve the efficiency of their macOS automation workflows.