Can you trust an autonomous LLM on your machine?

Watch an agent attempt to break our container: Agent-Service

Agent-Service is a demonstrably safe sandbox that lets ChatGPT-style agents run on your computer without abusing your files.

Here is a real-time look at an agent instructed to find a way to break the container. This cast is time-delayed and reviewed.

View the code. Hiring? Contact me.
agent-service-demo β€” safe mode

From Vibe-Coding to Verified Safety

This project aims to apply the confidence of software engineering to the age of vibe-coding. We start by putting a gptel-agent in a Guix container for monitoring and further study.

Our research focuses on creating reproducible, verifiable safety mechanisms for autonomous AI agents. By combining containerization, strict capability controls, and behavior monitoring, we enable safe and efficient scaling of LLM agents and sub-agents.

Read the Documentation

Get Involved

Run the Agent

Download and run the sandboxed agent on your machine. Experience safe autonomous AI execution firsthand.

Download for Guix

Contribute

Help us improve the sandbox, add features, or run experiments. Open to collaborators.

View on GitLab