Should I trust the decision-model lane for a long multi-step task?

Not yet. The README calls it optional and experimental and points you at the vision lane instead. It is cheap and quick on short single-goal web pages, but can still stop early or scroll in circles, and the guards reduce that without closing it.

What does it actually take to get this running on my desk?

A flashed Raspberry Pi Pico over serial, the Python serial dependency, a router API key, and a GUI terminal holding a macOS screen-recording grant. Safari also needs JavaScript from Apple Events enabled. Without a board, cliclick mode moves your real pointer so you can watch the loop.

How does it avoid clicking through a checkout by accident?

Money labels such as Pay, Save, Activate, Confirm, Apply, Update, Submit and Delete always pause once for a human, whatever confidence the model returned. Below the confidence floor for the operation or the target, the step becomes a verify_stop instead of a click.

How is this different from driving the DOM with a browser driver?

It reads an element table and sends input as USB HID, so the OS sees ordinary mouse and keyboard events with no WebDriver flag. That is also why it reaches Electron, games or a BIOS screen, and why it needs hardware and screen permissions instead of a headless browser.

What hardware do I need to run it?

A Raspberry Pi Pico flashed as a USB-HID mouse and keyboard, a board that costs about $4, plus the machine running the Python loop. No Pico is needed to try the loop: --hands cliclick moves your real pointer instead.

Which models does it call?

The default vision lane uses z-ai/glm-5.3-flash for planning and bytedance/ui-tars-1.5-7b for grounding, while the experimental Jev fast lane uses typesafe/jev-1.13. Everything runs over OpenRouter by default.

What safety guards does it have?

A click on a control whose label reads Pay, Save, Activate, Confirm, Apply, Update, Submit, or Delete always pauses once for a human, whatever the model's confidence. A scroll guard also allows at most three scrolls in a row.

CommunityJev Demos & Experiments

Ghosthands

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.

See highlights

Open github.com

Stars: 56
Language: Python
License: MIT
Published by: Community
Author: affirmitv
Last verified: Oct 2026

affirmitv/ghosthands

Organization repository on GitHub

View on GitHub

Give an agent real hands and eyes: undetectable, DOM-independent GUI automation. A $4 USB-HID microcontroller for hands, a vision grounding model for eyes, any LLM for a brain.

Forks: 8
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Give an AI agent real hands and eyes. ghosthands drives a real computer screen the way a person does: it looks at the screen, decides what to do, and moves a real mouse and keyboard over USB HID — so the operating system cannot tell it from a human, and it works on any app, not just a browser DOM. The Jev fast lane is still optional and experimental; the vision lane remains the default. Two merged changes hardened the loop on short web tasks:

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

Tagscommunity jevbest

Works forJev for Classification Jev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

The fast lane answers operation and target together at about $0.0001 and 0.4 s a step, against $0.00093 and 8.7 s for a screenshot decision.
Hands are a flashed Raspberry Pi Pico over serial; a cliclick mode drives your real pointer if you have no board yet.
DONE is re-checked by a small text model, and scroll, dead-click and loading guards keep the loop from spinning.
Labels like Pay, Save, Submit and Delete always pause once for a human, whatever confidence the model reported.

Watch out

The decision-model lane is explicitly optional and experimental; the vision lane is the default and the one the README tells you to rely on. The fast lane can still stop early on the wrong page or loop on long, multi-step tasks, and its benchmark covers five short web tasks in Safari on one Mac, so nothing here speaks to native apps or longer runs. It needs macOS screen-recording permission, Safari with JavaScript from Apple Events enabled, a router API key and flashed hardware, and it takes over your real pointer. The undetectability claim is the project's own, not a measured guarantee. 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

celesto

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

jevk5

147

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

Jev_Star

StarCraft II research repo combining Jev action selection with optional Astra plans across full-game macro control and SMAC-Hard micromanagement, with videos, replays, and a paper.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

sc2musa/Jev_StarPython

robodiag-harness

Terminal harness that checks ROS 2 robots with deterministic health tests, capability descriptions, SQLite history, and optional Jev evidence routing.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

YueBit/robodiag-harnessPython

jev-ultrafast-mcp

MCP server where one browser_goal call hands a URL, a goal, and code-checked assertions to a server-side loop driven by a decision model over live page elements.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

jiawei686/jev-ultrafast-mcpPython

jev-style

Local Qwen3.5 decision models in 0.8B and 2B sizes with fitted temperatures, systemone-shaped server, playground, guard, MCP tools, and eval harness.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

lawrence3699/jev-stylePython

From the Jev guides

Original write-ups on what this entry is for.

Per-Tick Decisions in Games

Sep 28, 2026

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

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai