Docker released Docker Sandboxes, a product that runs AI coding agents such as Claude Code, Gemini CLI, GitHub Copilot CLI, Codex, OpenCode and Kiro inside disposable, isolated microVMs, according to Docker's product page.
Each sandbox keeps a hard security boundary from the host machine, isolating the agent's filesystem and network access, while still giving it a working dev environment where it can install packages, edit configuration files and run its own nested Docker containers, Docker said. The host itself stays untouched.
Docker pitched the product for agents running with permissive settings, such as Claude Code's dangerously skip permissions flag, that let an agent act without a human reviewing every step. "Agents do their best work when they have freedom," the product page says. "Sandboxes let them run fast without running wild, so speed and safety stop being a tradeoff." A companion feature, Docker AI Governance, lets organizations set network and filesystem rules that apply across every sandbox in use.
Docker Sandboxes is available now. It installs on macOS through Homebrew and on Windows through winget, and does not require Docker Desktop, according to the product page.
The launch lands as more coding agents default to running autonomously rather than asking permission for each action. A disposable, hard walled sandbox is the piece that lets a team say yes to that autonomy without betting the host machine on the agent's judgment.