Docker Sandboxes: microVMs to run Claude Code and other coding agents safely

2026-01-30 · Publisher : Docker · Domain : Cloud & DevOps

On 30 January 2026, Docker launches the next generation of Docker Sandboxes, with microVM isolation on macOS and Windows: disposable environments to let Claude Code, Codex CLI, Copilot CLI, Gemini CLI or Kiro run unattended.

### What changes On 30 January 2026, Docker launches the "next evolution" of **Docker Sandboxes**, introduced a few months earlier as an experimental preview: isolation moves to dedicated **microVMs**, available on macOS and Windows. Each coding agent runs in a disposable environment where it can install packages, change configuration or run containers without touching the host machine. ### Key facts - Each agent runs in a dedicated microVM, with its own kernel and Docker daemon; only the project workspace is mounted. - Docker says Sandboxes are the only solution it knows of that lets an agent build and run containers while staying isolated from the host Docker daemon. - Supported agents: Claude Code, Codex CLI, Copilot CLI, Gemini CLI and Kiro, with the same experience. - Network isolation through allow and deny lists; reset in seconds if an agent goes off the rails. - Announced next: Linux support, MCP gateway, exposing ports to the host and additional agents. ### Why it matters The main brake on coding-agent autonomy is the fear of letting a tool run commands with the developer's rights. Classic containers isolate poorly a workload that needs to run Docker itself; a microVM solves that case. For an IT services firm, it is a candidate guardrail to evaluate for teams using coding agents on client repositories, before delegating long tasks without validating each command. Agents can install system packages and run services inside the microVM without approval at each step. ### Takeaway Docker offers the microVM as the safety envelope for autonomous coding agents.

Tags : Conteneurs, Agents, IA, Sécurité