JevMade Sign in
← Back to experiments

Browser & computer use

ghosthands

A GUI agent that drives a real computer through USB HID after Jev chooses an operation and indexed screen element.

Bookmark: ghosthands Keep this in your collection.
Leave a noteWhat would you try with this? : ghosthands

Only you can see your notes.

Source screenshot of ghosthands
SOURCE SCREENSHOTFull screenshot ↗

What it does

Actions that commit something always pause for approval. When the accessibility-style element table is missing, the system switches to a vision lane instead of guessing coordinates.

How you can use it

To start, list the repetitive computer jobs you want done. Note the fields to fill and buttons to click in any desktop app. Mark actions that spend money or save permanent changes so a person checks them first. A developer must connect a small accessory board that acts like a physical keyboard and mouse.

Your developer will need an access key that connects the software to an online AI provider. The system reads screen buttons directly or looks at screen images to decide its next move. It requires special screen recording permissions, and it cannot promise that every automated click will land correctly.

Maker-reported (not independently measured by JevMade): Jev fast lane: 2,387 input tokens, 223 output tokens, $0.00010 and 0.38 s for one decision on an 18-control page, measured 2026-09-18 through OpenRouter (author-reported) · 9x cheaper and 23x faster per decision than the vision lane (author-reported) · Whole live task: nine decisions, 10 seconds wall clock, $0.0015 total (author-reported) · Default confidence gates GH_JEV_MIN_CONFIDENCE 0.35 and GH_JEV_MIN_TARGET_CONFIDENCE 0.15 (source defaults)

Primitives
choice
Platform
Python
Added
Project created
GitHub stars
49 (snapshot, not live)

Keep this for later

Sign in to bookmark experiments, guides and videos, and keep notes only you can see.

Continue to sign in

We’ll bring you back to this listing.