AgentBox is an open-source command-line tool whose goal is to run multiple coding AI agents in parallel in sandbox virtual machines with a single command. The agents include Claude Code, Codex, and OpenCode. It can run on a local PC (Docker), a self-hosted server, or cloud platforms (Hetzner, Daytona, Vercel, E2B, DigitalOcean). Each box is an independent VM/container, equipped with a dedicated browser, screen sharing (noVNC), a persistent shell, and a prewarmed VS Code / Cursor IDE.
Main capabilities:
- Teleport: migrate a project to a dedicated VM with a single command, either locally or in the cloud.
- Automation: automatically bring in the agent's skills, plugins, and settings.
- Complete computer environment: each box has a browser, screen sharing, a persistent shell, and a prewarmed VS Code / Cursor IDE.
- Checkpoints: start a new box from a previous checkpoint in sub-second time; automatically pause when unused to save cost and resources.
- Security: git credentials remain on the local machine, and pushing to a remote repository requires a permission request.
- Terminal integration: can be used with iterm2, cmux, tmux, and Herdr.
- Cloud provider support: optionally configure Hetzner, Daytona, Vercel, E2B, and DigitalOcean, or run remote Docker on your own machine via SSH.
- Custom provider plugins: publish your own provider via @madarco/agentbox-provider-sdk.
Usage: Installation requirements are macOS (arm64 or Intel) or Linux, Docker (Docker Desktop or OrbStack), and Node >=20.10. The first agentbox create / agentbox claude will build an approximately 1 GB agentbox/box:dev image (one-time). Cloud setup is optional; use the agentbox install interactive wizard to configure providers. Common commands include: agentbox create to create a box, agentbox claude to create a sandbox and start Claude Code, agentbox url to open the web app URL, agentbox screen to open the VNC viewer, agentbox code to open in VS Code or Cursor via Dev Containers, agentbox shell to open an interactive bash, agentbox checkpoint to manage checkpoints, agentbox pause/unpause to freeze/resume a box, agentbox download to download the box workspace back to local, agentbox self-update to update itself, etc. For more details, see the official documentation at agent-box.sh/docs.
Use cases: developers who need to run multiple coding AI agents in parallel; those who want to isolate projects in sandbox VMs while retaining local git credentials; those who need to flexibly choose the runtime environment among local Docker, a self-hosted server, or cloud platforms.
Limitations: first use requires building an approximately 1 GB image; cloud providers require obtaining and configuring your own API token; the nightly channel requires joining via npm -g install @madarco/agentbox@nightly; security vulnerabilities must be reported via SECURITY.md; first-time contributors must sign a one-line CLA. Other cost, pricing, and specific performance results information is not provided.
