Microsandbox is a microVM runtime and library that executes untrusted workloads in fast, isolated virtual machines.
The tool addresses the need to safely run untrusted code—such as AI agents, user-submitted scripts, plugins, and automation tasks—without risking the host system. It achieves this through hardware-level isolation using microVM technology while maintaining the speed and convenience developers expect. Microsandbox boots microVMs in under 100 milliseconds on average and can be embedded directly into applications without requiring a separate daemon or server setup. It supports standard OCI container images from registries like Docker Hub and GHCR, and provides familiar Docker-like workflows for images, commands, shells, and volumes. The tool includes a mechanism for passing secrets into sandboxes in a way that prevents them from being leaked or exploited within the VM.
Microsandbox suits projects that need to isolate untrusted or user-provided code while maintaining fast startup times and local execution. It works across Linux, macOS, and Windows. The embeddable nature of the library makes it particularly valuable for applications that spawn sandboxes programmatically rather than managing them through external infrastructure. Long-running sandboxes can operate in detached mode for persistent sessions. The project also provides integrations for AI agents through companion tools that allow agents to create and manage their own sandboxes.
Development activity shows consistent engagement with the project. The team maintains active communication channels and continues to expand the ecosystem with complementary tools and integrations. Work spans both the core runtime and agent-focused extensions, indicating sustained investment in the platform's capabilities.
