jev-use: open-source Mac voice control reads the Accessibility tree, no screenshots, claimed 100x faster than LLMs

_AustinCalvert_ · x · 2026-09-21

Developer savka777 open-sourced jev-use, a macOS voice/typed computer-use harness built on TypeSafe's Jev model (65 stars on GitHub so far).

The key idea is no screenshots and no vision model: it reads the screen through the Accessibility tree, Jev picks the action and target element in one pass, macOS natively executes it, then the loop checks what changed and repeats. The author claims skipping vision makes it 100x faster than screenshot-based LLM approaches.

Usage: hold Control–Option–Space and speak, or type/CLI commands (scripts/say.sh "Open Finder"). Requires macOS 14.2+ and Xcode, zero dependencies, API key stored in Keychain.

Original post →

More from coding & agent

coding & agent channel →