- Published on
Docker Sandboxes – Disposable, isolated sandboxes for AI agents
- Authors
- Name
- DeepHorus
Docker has launched Sandboxes, a new product designed to provide disposable, isolated environments for AI coding agents, enabling safe autonomous operation. The tool allows developers to run agents like Claude Code and Gemini CLI in microVM-isolated sandboxes that protect the host filesystem and network. It is available immediately via Homebrew with the commands 'brew trust docker/tap && brew install docker/tap/sbx'.
The rise of AI coding agents has introduced new security challenges, as these agents often require broad permissions to complete tasks. Unrestricted access can lead to accidental data loss or security vulnerabilities. Docker's Sandboxes address this by offering a hard security boundary without the overhead of traditional VMs. The company describes the product as giving agents the autonomy they need to get work done, safely.
Available now via Homebrew, Docker Sandboxes supports leading coding agents including Claude Code, Gemini CLI, Copilot CLI, Codex, OpenCode, and Kiro. The sandboxes provide network and filesystem controls that users can define, are disposable by default, and spin up faster than VMs. Additionally, agents can spin up Docker containers within the sandbox, creating a realistic development environment where they can install packages, run services, and work unattended.
This development matters for organizations deploying AI agents at scale, as it enables YOLO mode – which skips permission prompts for speed – to be used safely. By isolating each agent inside a dedicated microVM, Docker claims to eliminate the risks typically associated with autonomous agent execution. The company emphasizes that sandboxes are fully isolated, offering more isolation without the full cost of running a VM.
The tech community has long called for better guardrails on AI agent operations, and Docker's approach appears to be a response to that demand. Notably, the product does not require Docker Desktop, making it accessible to a broader range of developers. For teams needing centralized controls, Docker AI Governance extends the functionality with network access policies, filesystem rules, and org-wide MCP governance.
Looking ahead, Docker plans to expand control through AI Governance, allowing administrators to define policies once and enforce them across every developer's machine. For now, Sandboxes offer a practical solution for individual developers and teams seeking a safer way to grant AI agents autonomy. With support for major coding agents and a lightweight, fast-spinning architecture, the product is poised to become a standard tool for AI-assisted development workflows.
