What screen data actually leaves my Mac during a run?

Pixels, screenshots, and coordinates stay on the Mac. Only observed text, the goal, and finite choices go to the selected provider, and guarded input rechecks focus, occlusion, fresh frames, and live accessibility state.

How do I configure Vercel, OpenRouter, or direct keys?

Copy the example env file, set the provider selector, and add only that provider token. A missing or rejected token stops the run, Vercel and OpenRouter have live checks, and the direct adapter still needs verification with your key.

What setup is required before the first automation?

You need Apple silicon, recent macOS, Xcode command line tools, modern Node, and uv for one-time model conversion. Setup downloads a pinned detector, verifies its hash, and converts it to CoreML.

Are benchmark passes just Jev grading its own work?

No. Each benchmark uses disposable apps with a separate result log the policy cannot read. The Finder demo also uses a separate verifier for file locations and content hashes, independent of the DONE decision.

What hardware and software does it require?

Apple silicon, macOS 14.2 or newer, Xcode Command Line Tools, Node.js 22 or newer, and uv for the one-time CoreML model conversion. The setup downloads a pinned OmniParser detector, verifies its SHA-256 and converts it; model weights, credentials and local captures are gitignored.

How do the three provider routes differ?

JEV_PROVIDER selects vercel, the default with AI_GATEWAY_API_KEY and model typesafe-ai/jev, openrouter with OPENROUTER_API_KEY and ~typesafe/jev-latest, or typesafe with TYPESAFE_API_KEY and jev-latest. A missing or rejected token stops the run, and the runner never silently switches providers.

What does the Finder batch demo actually do?

It creates nine dummy files and three empty folders, has Jev classify all nine filenames in one request, then selects each group through macOS accessibility and moves it with one native drag, three group moves in total. A separate verifier checks all nine locations and content hashes.

CommunityJev Demos & Experiments

Jev Macos Loop

Native macOS computer-use agent on Apple silicon that finds controls with local CoreML detection, OCR, and accessibility, then asks Jev to pick the next finite-choice action.

See highlights

Open github.com

Stars: 24
Language: JavaScript
License: AGPL-3.0
Published by: Community
Author: jcpsimmons
Last verified: Oct 2026

jcpsimmons/jev-macos-loop

User repository on GitHub

View on GitHub

Open-source macOS AI computer use and native GUI automation on Apple silicon. Jev + OmniParser CoreML + Apple Vision OCR. Bring your own OpenRouter, Vercel AI Gateway, or TypesafeAI token.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Real Finder demo: 9 files → 3 group moves in 7.39 seconds, independently verified. Watch the full-resolution MP4 · Recording details. Jev macOS Loop is an open-source computer-use agent for native macOS GUI automation on Apple silicon. It combines OmniParser CoreML, Apple Vision OCR, and macOS accessibility to identify controls locally, then uses Jev to select the next action. Bring your own token for Vercel AI Gateway, OpenRouter, or TypesafeAI.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import

Works forJev for Classification Jev for Routing

Used inJev for Software & DevTools

Languages: JavaScript

Highlights

Combines ScreenCaptureKit frames, OmniParser CoreML, Apple Vision OCR, and accessibility labels entirely on-device.
Sends text-only finite-choice decisions to your own Vercel, OpenRouter, or TypesafeAI token with no silent fallback.
Includes a Finder batch demo that classifies nine files in one request and moves three groups with native drags.
Reports sub-second observe-decide-act cycles on small native suites, with traces and failure notes included.

Watch out

Runs are limited to disposable test apps and small functional samples, timings exclude warmup, and tokens never switch providers silently. 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.

More Jev resources like this

jev-by-example

Ten runnable JavaScript experiments on agent decisions, each pairing a small Jev judgment with visible application policy code, plus authored fixtures so the whole set runs offline.