A macOS computer-use agent replaces the vision model with OCR plus a TypeSafe classifier and claims $0.0002 per step

GitHub (awlevin/typesafe-computer-use)·high signalXBlueskyLinkedInCopy link

typesafe-computer-use (created 2026-09-16, 128 stars) screenshots the frontmost window, OCRs it deterministically, merges the text blocks with accessibility-tree controls into one numbered list, then asks TypeSafe to pick among three mutually exclusive next actions rather than having a large vision model reason over pixels. Measured decision latency is 0.13 to 0.38 seconds against roughly 5.2 seconds for Claude Opus, with end-to-end steps at ~1.5s versus ~5.5s. The cost claim, ~$0.0002 per step against ~$0.032, rests on TypeSafe returning probability distributions over choices without charging for generated tokens.