Hanzo Computer
An agent that uses the computer in front of it
It reads pixels
No DOM, no accessibility tree, no selectors to maintain. A native app, a canvas, a remote desktop and a web page all look the same to it — like a screenshot.
OCR when clicking is hard
Add -with-ocr to the model and it reads text off the screen to build a map of clickable elements — the fix for a button a model can see but keeps missing by twenty pixels.
Set-of-Mark prompting
Add -with-som and the screenshot arrives pre-labelled, so the model names a marked element instead of guessing a coordinate.
-with-ocr · -with-som
Mouse and keyboard
Mac, Windows, and Linux with an X server. On macOS the terminal asks for Screen Recording and Accessibility the first time, because those are exactly the permissions this needs.
Say it out loud
operate --voice takes the objective by microphone. Needs the audio extras and portaudio installed; it is not in the base install.
operate --voice
It has whatever you have
This runs as you, on your machine, with your logged-in sessions. Give it a fresh user account or a VM before you give it anything that matters — and remember that text on a web page can try to redirect it.
Paste and ship
Paste this into any agent. It reads the skill manifest and calls Computer from there.
Read https://hanzo.ai/skill.md and use Hanzo Computer in my project. Start with: hanzo desktopFor the software with no API
The vendor portal that exports one CSV at a time. The desktop tool from 2009. The internal app nobody will reopen. Those.