Docker Sandboxes – Disposable, isolated sandboxes for AI agents

Docker Sandboxes provide disposable, isolated environments for AI agents. These sandboxes allow agents to spin up containers, install packages, run services, and operate unattended. Popular AI agents like Claude Code, Copilot CLI, Codex, Kiro, and OpenCode can operate within these microVM isolated environments, which safeguard the underlying filesystem and network from potential agent-induced harm.

Out of the box, Docker Sandboxes support a range of AI agents, including Claude Code, Gemini CLI, Copilot CLI, Codex, OpenCode, and Kiro. For those who wish to grant agents greater autonomy, there's a YOLO mode (--dangerously-skip-permissions) that removes approval prompts. However, this mode should be used with caution due to the increased risk involved.

The Docker Sandboxes approach ensures safety by isolating each agent within a dedicated microVM. This level of isolation offers more security than running a full virtual machine, while still allowing agents to perform tasks that require elevated permissions, such as running additional Docker containers. To maintain consistency and enforce these controls across a team, Docker AI Governance offers centralized management solutions.

This includes managing network policies, filesystem rules, and MCP governance, enabling administrators to define these once and ensure their enforcement on every developer's machine.