Docker Sandboxes: Isolated microVM environments for AI coding agents
Original: Docker Sandboxes – Disposable, isolated sandboxes for AI agents
Why This Matters
Isolated agent execution environments address a critical infrastructure gap as AI coding agents become standard developer tools.
Docker has launched Docker Sandboxes, a product offering disposable, isolated microVM environments designed for AI coding agents such as Claude Code, Gemini CLI, GitHub Copilot CLI, Codex, OpenCode, and Kiro, enabling safe, unattended code execution without exposing the host system.
Docker has introduced Docker Sandboxes, a new product aimed at providing secure, isolated execution environments for AI coding agents. Each sandbox runs inside a dedicated microVM with the user's project workspace mounted in, allowing agents to install packages, modify configurations, and spin up their own Docker containers — all without affecting the host machine. The product supports leading coding agents including Claude Code, Gemini CLI, GitHub Copilot CLI, Codex, Kiro, and OpenCode. Installation is available via Homebrew on macOS (`brew trust docker/tap && brew install docker/tap/sbx`) and Winget on Windows (`winget install Docker.sbx`). Sandboxes are disposable by design, described as faster than traditional VMs to spin up and tear down. Docker emphasizes that permissive 'YOLO mode' (e.g., `--dangerously-skip-permissions`) can be used with confidence inside the sandbox boundary. Network and filesystem controls are user-defined and can be enforced organization-wide through Docker AI Governance, a separate product. The offering is positioned as eliminating the tradeoff between agent autonomy and security.