AgentBox

Canonical version: AgentBox.

AgentBox is Rui Carmo's Docker-based sandbox for running coding agents in isolation. The README is refreshingly honest about the state of the art: "There's no perfect way to sandbox agents (yet), but containers are a practical start." MIT licensed, ~141 stars, mostly Dockerfile and Makefile.

What it provides

• A Debian Trixie development container with common tooling: gh, Nushell, lazygit, Homebrew, APT, uv, Bun, Node, Go
• Preinstalled coding agents (Copilot CLI, Codex, Pi), with optional installs for OpenCode, Gemini CLI, and others
• Services off by default, enabled via environment variables: ENABLE_DOCKER (see Docker in docker (dind)), ENABLE_SSH, ENABLE_RDP
• A workspace skeleton at /home/agent/workspace-skel and webterm integration for multi-session management
• Roadmap: CPU/memory limits, network isolation, and stronger isolation via gVisor or Kata Containers

Why it matters

Running agents with --dangerously-skip-permissions on a bare machine is a bad idea; a disposable container makes the YOLO workflow defensible. AgentBox packages that pattern so you don't have to build your own Dockerfile for every agent. Worth comparing with Docker Sandboxes, which attacks the same problem from the official Docker side.

The candid framing is the useful part: container isolation is a practical baseline, not a security boundary you should bet the house on. Anything truly sensitive stays out of the sandbox.

References

• Repository: https://github.com/rcarmo/agentbox

Related

• Docker
• Docker Sandboxes
• Docker in docker (dind)
• AI Agents
• Claude Code
• OpenCode
• Gemini CLI
• GitHub Copilot CLI

About Sébastien

Ready to get to the next level?