madarco/agentbox

What it solves

AgentBox provides a way to run multiple AI coding agents (such as Claude Code, Codex, or Open Code) in parallel within isolated, sandboxed environments. It eliminates the friction of manually setting up virtual machines or containers for each agent, allowing developers to "teleport" their project into a dedicated environment with all necessary settings and tools pre-configured.

How it works

AgentBox uses Docker containers (with FUSE overlays) to create "boxes"—dedicated virtual environments that can be hosted locally or on various cloud providers (e.g., Hetzner, Vercel, E2B, DigitalOcean). Each box includes a persistent shell, a browser, and integration with IDEs like VS Code or Cursor. Users can launch agents via a simple CLI command, manage them through a dashboard, and use checkpoints to start new boxes from a saved state in under a second.

Who it’s for

It is designed for developers who use AI coding agents and want to run multiple agents simultaneously on different tasks without cluttering their local machine or risking their primary environment.

Highlights

• Multi-Agent Parallelism: Run several agents in separate boxes simultaneously with a single command.
• Cloud & Local Flexibility: Supports local Docker as well as multiple cloud providers for hosting sandboxes.
• Instant State Recovery: Sub-second startup of new boxes using checkpoints.
• Integrated Tooling: Built-in support for VNC screen sharing, web app URL tunneling, and direct IDE connection via Dev Containers.
• Secure Credential Handling: Keeps git credentials on the local machine, requesting permission before pushing to remote repositories.