dev-sandbox

dev-sandbox is a versatile tool designed to run AI coding agents within isolated Podman containers, optionally enhanced with krun microVMs. This project focuses on providing an environment that maintains the security and integrity of your system by ensuring that untrusted code runs in complete isolation, mitigating various security risks associated with executing arbitrary code.

Developed for Linux distributions compatible with Podman, and specifically tested on Fedora 44, dev-sandbox simplifies the process of building, configuring, and executing containers through a single self-contained bash script. No additional dependencies are required beyond Podman, allowing for ease of installation and use.

Key Features

• Isolated Environments: Each agent operates in its own dedicated space, with only the current project directory accessible, thereby preventing unauthorized access to sensitive files such as SSH keys, browser cookies, and API keys.
• Risk Mitigation: By isolating agents, dev-sandbox helps to guard against data exfiltration from malicious packages and prompts that could execute harmful commands.
• Multi-Profile Support: Users can choose from multiple predefined profiles, such as the default claude for Claude Code, or create custom profiles for specific needs, ensuring clear categorization and management of agents.
• Flexible Networking Options: Various networking configurations allow users to control traffic flow, enforce strict network rules, and run applications in environments that require specific connection setups.

Usage Example

# Run an AI coding agent in a secured environment:
dev-sandbox                         # Starts with the default claude profile

Profiles and Agent Management

Profiles include:

• claude: For Claude Code, characterized by a green prompt.
• research: A template for adding custom agents, marked with a red prompt.
• agy: Optimized for Google AI, distinguished by a yellow prompt.
• vncgui: For running GUI applications with a purple prompt, allowing VNC+XFCE desktop environments.

Users can quickly switch between profiles and leverage built-in commands to manage agents effectively or extract configuration details for specific profiles.

Custom Configuration

The configuration system allows users to define defaults, override settings for each profile, and even add new profiles tailored to their specific requirements. Environment variables can be passed securely from the host to ensure sensitive information is handled safely.

Overall, dev-sandbox stands out as an essential utility for developers needing to run AI coding agents in a safe and controlled manner. For extended use cases and technical architecture details, visit the documentation and explore its capabilities further.