User repository on GitHub
macOS computer use driven by Jev (TypeSafe System One) as the decision maker
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
Computer Use Jev
Go driver for native macOS apps where Jev picks the next Accessibility action from live app and window state, with code mapping typed answers to tool calls.
- Stars
- 6
- Language
- Go
- License
- MIT
- Published by
- Community
- Author
- paulsmith
- Last verified
- Oct 2026
About this project
Drive macOS applications from Go — with Jev (TypeSafe's System One decision model) as the decision maker. Extracted from herbie's computeruse tool and extended with a typesafe decision layer. computeruse — operates native macOS applications through the Accessibility API via a persistent Swift worker (computeruse/swift/). Actions: apps, windows, activate, snapshot, click, fill, type, press, screenshot. Tokens (a1, w2, e5) reference apps, windows, and elements discovered at runtime.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Go
Highlights
- Persistent Swift worker drives apps, windows, activation, clicks, fills, typing, presses, and screenshots.
- Low-confidence choices stop the run and report the distribution instead of acting blindly.
- Targets come only from tokens present in the current UI listing, keeping closed sets closed.
- Single instructions run on Jev alone; multi-step goals are split once by a configured LLM then stepped by Jev.
Watch out
Requires macOS 14 or later, Accessibility and Screen Recording permissions, Xcode Command Line Tools, and TYPESAFE_API_KEY. Open currently means activating an already running app with no launch action, text stays literal, and a valid LLM split is not proof its semantics are correct. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
The LLM never gets desktop tools. Each step gathers app list, windows, and last UI state, Jev answers typed questions about action, target token, completion, and text need, and code maps those answers to a tool call.
If action confidence drops below the threshold, the run stops and reports the distribution rather than acting blindly. Unknown or low-confidence semantic shortcut choices are also not sent.
Jev first judges single versus compound. A single instruction passes through unchanged with no LLM config; compound goals are split once by an OpenAI, Anthropic, or OpenRouter model into ordered JSON, then run through Jev in the same worker session with a total max-steps budget.
Grant Accessibility and Screen Recording to the terminal, build the Swift worker once with Xcode tools, keep target apps like TextEdit already running, quote literal text with double quotes inside single-quoted shell args, and preview consequential plans with dry-run.
A persistent Swift worker drives native macOS apps through the Accessibility API, with actions apps, windows, activate, snapshot, click, fill, type, press, and screenshot. Runtime tokens such as a1, w2, and e5 reference apps, windows, and elements discovered in the current snapshot.
The splitting LLM supports the OpenAI Responses protocol (openai), Anthropic Messages (anthropic), and OpenAI Chat Completions (openrouter), selected with COMPUTER_USE_JEV_PROVIDER and an explicit COMPUTER_USE_JEV_MODEL. No provider or model is chosen silently.
It activates an already running application; the tool does not add an application-launch action, so the target app must already be running. For example, TextEdit has to be open before a goal that types into it will work.
More Jev resources like this
2GitHub stars
Browser-use loop where a ~200 ms Jev call picks every step, run twice over the same site and model to isolate what a site map changes.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
sightmap/jev-turboGo
sightmap/jev-turboGo46GitHub stars
Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.
Jev Demos & Experiments#community#jevbest#catalog-import
941GitHub stars
Library that adds schema-guaranteed typed outputs to existing autoregressive LLMs without changing weights, served through SGLang or a hosted API.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
TypeLLM/TypeLLMPython
TypeLLM/TypeLLMPython98GitHub stars
Self-hosted Docker service and web UI that packages Laya structured-decision inference behind a compatible System One endpoint.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
1Panel-dev/laya-serverPython
1Panel-dev/laya-serverPython39GitHub stars
Mac menu-bar voice controller where Jev picks the next action from observed Accessibility controls and Swift executes it in an observe-decide-execute loop.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
ronadin2002/jev-cuaSwift
ronadin2002/jev-cuaSwift55GitHub stars
Toolkit that makes a model behind an OpenAI-compatible endpoint or running in-process answer typed choice, score, and yes/no questions from next-token probabilities in Jev API shape.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
fidecastro/jevifyPython
fidecastro/jevifyPythonFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.