microsandbox: Isolate AI agent code with 100ms boot times

Useful when you need to safely run untrusted workloads such as AI agents, user code, CI jobs, and scrapers in a local environment. Unlike traditional containers, it ensures isolation by leveraging hardware-level microVM technology, providing fast boot times averaging under 100ms. You can create VMs directly within your code without server configuration or daemon execution, making it suitable for embedded environments.

It natively supports container images from standard OCI registries like Docker Hub and GHCR. It offers a workflow similar to Docker for managing images, commands, shells, and volumes, but is designed to prevent secret keys from leaking into the VM, ensuring high security. It runs on Linux, macOS (Apple Silicon), and Windows (with WHP support), and allows you to control the sandbox lifecycle via SDKs for various languages including Rust, Python, TypeScript, Go, and Ruby.

It supports Agent Skills and MCP servers, enabling AI agents to create and manage sandboxes autonomously. It integrates with major AI coding agents like Claude Code, Cursor, and GitHub Copilot to automatically configure isolated workspaces. It is applicable to various real-world scenarios, such as running Docker inside a sandbox, performing Playwright headless browser tasks, and rehearsing database migrations.

Currently in beta, there may be breaking changes or unsupported features. It has already been adopted and used by several projects, including Vercel's Eve and LlamaIndex's sandboxed-lit. It is suitable for developers who need an isolated execution environment in their local development setup or who need to safely control code execution by AI agents.