New macOS Agent Cuts Decision Costs 155x Over Claude

Developer Andrew Levin has introduced typesafe-computer-use, an MIT-licensed, open-source desktop agent designed for macOS 14 and higher. Running on Python 3.12 or later, the tool aims to solve the high latency and cost associated with traditional frontier-model agents. Instead of sending full screenshots to a large model for every step, this agent processes screen state locally to make decisions costing approximately $0.0002 each.

The system achieves its efficiency by keeping screenshot pixels on the local machine. It extracts local data using optical character recognition, accessibility metadata, active application details, and browser URLs. This structured state is sent to TypeSafe, a hosted decision model that serves as the project's core dependency. TypeSafe can evaluate up to 255 options simultaneously, returning calibrated choice probabilities in milliseconds with zero output-token charges. For tasks requiring free-form text generation, such as filling out forms or writing URLs, the agent routes requests to a separate writer model, which defaults to Claude Haiku.

According to repository benchmarks, this architecture makes decisions 155 times cheaper and 14 to 40 times faster than Claude Opus 5, with a reported decision latency of just 0.13 to 0.38 seconds. However, these figures represent a narrow point comparison and have not been independently verified. The metrics do not account for overall task-completion rates, recovery benchmarks, or the added latency and cost of the writer-model calls.

For AI developers and practitioners, typesafe-computer-use offers a highly cost-effective framework for building local automation tools without compromising data privacy. By avoiding the transmission of raw screen images to external APIs, it addresses major enterprise security concerns. Additionally, the inclusion of replayable run folders simplifies debugging when the agent stalls, providing a practical environment for refining automated workflows.