User repository on GitHub
Superfast action layer for computer-use agents, on MacOS.
- Forks
- 26
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Demos & Experiments
Arc Cua
A macOS action layer that lets a planner hand a bounded desktop subtask to a fast decision model which runs the UI loop, plus a driver you can run on its own. Jev is the default backend but the transport is pluggable, and the planner lives outside the package.
- Stars
- 314
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- shhivv
- Last verified
- Oct 2026
About this project
This repository ships two separate tools in one package: If you came for the driver, you need nothing below this section. arc-driver does not use decision models, JEV or TypeSafe; it is a standalone driver.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- The decision model picks only from an action space rebuilt from the current desktop each step, so it cannot target elements that are not present.
- Controls whose labels read like consequential actions are withheld unless the subtask allows that category, and a risky click returns NEEDS_AGENT.
- Secret inputs show as a placeholder everywhere the model reads them while the runtime types the real value; results, output and logs are redacted.
- The macOS driver acts on windows in the background through accessibility APIs, so the pointer, front app and windows stay as the user left them.
Watch out
This is macOS-only and leans on Apple accessibility permissions, which is the main operational constraint; the driver installs on its own from any MCP client with arc-cua mcp, but the decision layer needs a provider behind it. A provider that answers with a choice outside what was offered executes nothing — the policy re-asks once by default and then raises, so a flaky model surfaces as failures rather than wrong clicks. Two gaps are stated rather than hidden: screenshots sent for screenshot checks or steps are not redacted, so secret inputs do not cover image evidence, and every literal string must come from the agent or the run stops with NEEDS_INPUT. No accuracy, latency or cost measurements are published, and the accessibility-plus-OCR perception and freshness guard are described in outline rather than benchmarked. 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.
No, and that is the central safety property. Text fields are offered only so the model can ask for a value it does not have; if nothing in inputs supplies one, the run stops with NEEDS_INPUT and returns the field's element id, role, name, current value and dropdown options so the agent can supply it and retry. Targets are restricted to elements observed on the current desktop, and hotkeys are validated against the defaults plus the subtask's own declared shortcuts, including for custom policies.
Controls whose labels contain whole-word matches for the delete, send, purchase or close categories are not offered to the decision model at all unless the subtask opts in through allowed_risks. That covers click and double-click, and the runtime refuses such a click from any policy, returning NEEDS_AGENT with the reason and not acting. It is a label-word heuristic though, so a destructive button that avoids those words is not covered by anything documented here.
Neither. The loop talks to a ChoicePolicy that builds the finite-choice questions and validates the answers, and a ChoiceTransport that ships them to a provider; TypeSafeJevPolicy is simply ChoicePolicy with the TypeSafe transport, and any provider returning a choice, confidence and probabilities fits. On the planning side, any GPT, Claude, Gemini, local or deterministic planner builds the payload — the package intentionally ships without one.
Treat it as the loop's exit signal, not a verified outcome. You supply the verification criteria in the payload yourself, and the four statuses are SUBTASK_COMPLETE, BLOCKED, NEEDS_AGENT and NEEDS_INPUT, so the calling agent has to decide what each one means. There are no published success rates, latencies or cost comparisons against a frontier model in the loop. The argument for the design is fewer expensive reasoning calls per task, which the README does not quantify.
arc-driver is a macOS driver that lets your own agent read and act on an app's window in the background and needs no model or API key. arc-cua is the decision-model action loop that hands a bounded subtask to a fast decision model; only it needs a model or key. A driver-only user needs nothing else in the README.
Literal values always originate from the agent through inputs; Jev never invents arbitrary text. When a field needs a value no input provides, the run stops with NEEDS_INPUT and returns the field's element_id, role, name, current value, and a dropdown's options so the agent can add the value and rerun the subtask.
allowed_risks is an array from delete, send, purchase, and close, defaulting to empty, so those action classes are gated unless declared. secret_inputs lists inputs keys whose values the model never sees, letting the agent supply sensitive literals without exposing them to the decision model.
More Jev resources like this
1kGitHub stars
Sandboxed virtual machines for agents: each sandbox is a microVM that starts in about 500 ms, keeps files and state between sessions, and runs locally or in Celesto Cloud behind a CLI and Python SDK.
Jev Demos & Experiments#community#jevbest
Community
CelestoAI/celestoPython
CelestoAI/celestoPython70GitHub stars
Research toolkit that turns embedding models into Choice, Score and Noul decision models, covering data synthesis, LoRA fine-tuning, serving, benchmarking and a playground.
Jev Demos & Experiments#community#jevbest
Community
HITsz-TMG/JevEmbedPython
HITsz-TMG/JevEmbedPython65GitHub stars
Non-commercial research harness that turns a branching narrative into decision-tree JSON so any OpenAI-compatible model, Claude Code or Jev can play it node by node.
Jev Demos & Experiments#community#jevbest
Community
Baba88611/detroit-ai-playerPython
Baba88611/detroit-ai-playerPython147GitHub stars
Apache-2.0 open-weight stand-in for a hosted decision model: Qwen3.5-4B with a distilled adapter scores noul, choice and score questions through a systemone-shaped server, with no generated tokens.
Jev Demos & Experiments#community#jevbest
Community
allebee/jevk5Python
allebee/jevk5Python56GitHub stars
macOS agent loop that moves a real USB HID mouse and keyboard under a vision model, with an experimental decision-model lane picking the next operation and screen target from an element table.
Jev Demos & Experiments#community#jevbest
Community
affirmitv/ghosthandsPython
affirmitv/ghosthandsPython46GitHub 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
From 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.