Harnesses that write and run code with your uncensored models, each in its own VM, with your keys kept outside and your files back only after you review them.
Its changes come back only after you approve them.
One command to a sealed VM.
Pick any of 21 harnesses for coding or security testing, answer three questions, and it starts in its own Sandbox VM.
yougori inference run HARNESS --server https://acme.yougori.com
Starting Yougori engine
Reading this computer's capacity
What should the harness work on?
This folder: C:\work\shop
What can the harness reach?
Only this inference
Prompts and replies always go to acme.yougori.com. Yougori relays them and keeps the API key outside the VM.
Size the harness VM
2 CPU, 4 GB memory
Start HARNESS with these settings?
Start
Starting Yougori Sandbox VM. It reaches only this inference: Yougori relays its model requests.
Preparing the tool inside your VM...
Your original project is unchanged. Review the VM copy with:
yougori sandbox review exe-4f2a91
Nothing lands until you approve it.
The harness works on a copy of your folder. You look at each change and write only the files you approve.
Review changes
src/cart.ts
+18 −4
src/cart.test.ts
+42
package.json
+1 −1
scripts/deploy.sh
+6 −2
Show the change
Promote 2 approved files
Discard
Deletions never apply
A file the harness deleted stays in your folder.
Originals backed up
Every file it replaces is kept first.
Late changes refused
A file that changed after you reviewed it is not written.
What never enters the VM.
Each harness gets its own kernel and a copy of your folder. Your secrets stay on the outside.
Your API key
Your real folder
Your home folder
Your SSH agent
Your provider accounts
Your Docker socket
Your computer
Keeps the key and adds it to each model request
Sandbox VM
Its own pinned kernel
Harness
No root. No network card unless you allow the whole internet
One link
Owner-only and token-authenticated, for the terminal, model requests and the sites you allow
Questions.
What is a harness?
An AI agent that reads, writes and runs code for you, with a model behind it. Yougori runs each one in a sandbox VM connected to your inference.
What makes it uncensored?
It works with the models on your own inference, uncensored ones included. No provider sits between the harness and the model to refuse a request.
Is it safe to run?
Each harness runs in its own VM with its own kernel, no root and no network card unless you allow the whole internet, through one authenticated link. Your keys stay outside the VM, and files come back only as you review them.
What can it reach?
You choose before it starts: only your inference, your inference and the sites you name, or the whole internet as well. Prompts and replies always go to your inference.
Which harnesses can we use?
The harnesses on your inference website, and any harness or tool that speaks the OpenAI API.