savka777/jev-use
Say it, and your Mac does it. A computer-use harness on Jev that reads the screen through Accessibility. Fast, no vision model
About savka777/jev-use
savka777/jev-use is an open-source project on GitHub, mainly written in Swift. Say it, and your Mac does it. A computer-use harness on Jev that reads the screen through Accessibility. Fast, no vision model It currently holds 71 stars and 7 forks with 1 open issues, and was last pushed on 2026-09-19 (repository created 2026-09-19).
Project Overview
AI Homed tracks it on the Today's Trending board, currently at rank #97 with 0 new stars today.
GitHub Repository Details
README

jev-use
Voice and typed computer use for macOS. You say what you want. Jev picks the next on-screen action. macOS performs it. No screenshots: the app reads the screen through the Accessibility tree.
Quick start
Requires macOS 14.2+ and Xcode. No dependencies.
bash build.sh
open "$HOME/Applications/Desktop Voice.app"
In setup: save your TypeSafe API key (stored in the Keychain), then allow Accessibility, microphone and speech.
Hold Control–Option–Space, speak, release. Escape cancels. Or type a command in the widget, or from a shell: scripts/say.sh "Open Finder".
Examples
- "Open Obsidian, create a new note and type hello"
- "Go to youtube.com, search Rick Astley and play the first video"
- "Open 3 new tabs"
- "Scroll down three times"
- "Tile all the Brave windows so none are stacked"
- "In every Brave window, go to wikipedia.org and search for accessibility" — Jev works out the steps once, code repeats them in each window
- "Close the window", "Save", "New tab" — any item in the app's menu bar
How it works
One loop, about 0.3–1.5 s per step:
1. Read. Walk the front app's Accessibility tree (~120 ms). Every element describes itself: what it is, its name, its value, where it sits, what it can do. No per-app code.
2. Choose. One request to Jev (jev-latest): the goal, the numbered targets, the last ten actions and their effects. Jev selects an operation and a target. It never generates free text; typed text is a span of your sentence.
3. Act. Press, select, type, menu, key, scroll, open, arrange windows.
4. Check. Read the screen again. Report the real effect. Repeat until DONE, BLOCKED or WAIT.
Low-confidence and destructive picks stop and ask instead of acting.
What is sent
To https://api.typesafe.ai/v1/systemone: your command, the app and window names, the on-screen targets with their labels and values, and recent actions. Secure text fields are excluded. No screenshots. Speech uses Apple Speech.
Everything is logged locally: log show --predicate 'subsystem == "local.jev-use"' --last 10m --info
Develop
swift test # 10 tests
bash build.sh # quit the app first