jev-macos-loop

September 20, 2026 · View on GitHub

Native Apple silicon macOS computer-use agent: local OmniParser CoreML + Vision OCR + Accessibility identify controls; Jev chooses the next guarded GUI action. Screenshots stay on-device.

At a glance	Details
Source	Source
Maintainer	jcpsimmons. Independently curated; this entry is not an upstream submission or endorsement.
Format	Node.js (≥22) CLI and loop (`jev-macos-loop` 0.1.0); CoreML model export via `uv`/Python 3.12.
Requirements	Apple silicon, macOS 14.2+, Xcode CLT, Node.js 22+, `uv` for one-time OmniParser conversion. BYOK for Vercel AI Gateway, OpenRouter, or TypesafeAI (`JEV_PROVIDER`).
License	AGPL-3.0.

When to use

Use it to automate native Mac apps (Finder demos included) when you want local perception and text-only Jev decisions. Prefer typesafe-computer-use for a Python OCR/Accessibility study path, or browser agents such as Jev Ultrafast when the target is the web.

How it works

Local ScreenCaptureKit / OmniParser / Vision / Accessibility build a control inventory. src/providers.mjs routes typed evaluation to Vercel (typesafe-ai/jev), OpenRouter Decisions (~typesafe/jev-latest), or TypesafeAI https://api.typesafe.ai/v1/systemone (jev-latest). Only observed text, the goal, and finite choices leave the machine. Input is guarded (focus, occlusion, confidence) before clicks or scoped Finder drags. A separate verifier checks outcomes independently of Jev's DONE decision.

Get started

Requires a local Apple silicon Mac (not runnable on this Linux review host):

git clone https://github.com/jcpsimmons/jev-macos-loop.git
cd jev-macos-loop
git checkout 1aadc01ef262c3b01460909c5dd102f0d9616ba5
npm ci
uv venv --python 3.12
uv pip install --python .venv/bin/python -r scripts/model-requirements.txt
.venv/bin/python scripts/export_model.py
npm run build
cp .env.example .env.local   # set JEV_PROVIDER and the matching token
npm start -- --doctor

Live runs send decision text to the selected provider and can incur charges. This listing did not install CoreML models or call TypeSafe.

Examples and demos

Finder batch demo media and notes under docs/media/ and docs/recording.md.
docs/providers.md and npm run check:provider (three small decision requests; no screen access).
test/providers.test.mjs: offline provider contract tests.

Limits and data handling

AGPL-3.0 applies to the agent code. Model weights and credentials are gitignored. Tokens go only to the selected provider; the runner does not silently switch providers. Direct TypesafeAI adapter has offline contract tests; upstream notes live Typesafe verification still needs your console token. Accessibility and screen-recording permissions are required on macOS.

Review and maintenance

Reviewed on 2026-09-20 at commit 1aadc01: 0.1.0, AGPL-3.0. AI-assisted source review of providers.mjs, README, LICENSE, and provider docs. No macOS install, CoreML export, GUI automation, or live provider calls were run on the Linux review host.