Microsandbox
lastmod 2025-06-01created 2025-06-011 edits
Virtual machines that feel and behave like containers
About
- microsandbox provides safe execution of untrusted user and AI code with virtual-machine-level isolation
- Overcomes the drawbacks of existing VMs and containers with ultra-fast boot (under 200ms), OCI container compatibility, self-hosting, and more
- Maximizes integration efficiency for developers and AI tools with SDKs for many programming languages and a CLI tool
- Suited to a wide range of AI and development use cases, including code execution, development environments, data analysis, web automation, and app hosting
- All work can be managed per project, with support for system-wide installation and persistent-session or isolated execution environments
- microsandbox is an open-source, self-hosted platform designed for safely executing untrusted user code or AI code (e.g. AI agents, user-submitted code, experimental code)
- Existing options have drawbacks: local execution has security vulnerabilities, containers offer incomplete isolation due to the shared kernel, traditional VMs boot slowly, and the cloud lacks flexibility
- microsandbox supports true process isolation based on microVMs (ultra-lightweight virtual machines) while offering the same fast startup and developer-friendly experience as containers
- Differentiated by boot within 200ms after initial environment setup, container image (OCI) compatibility, MCP-based AI integration, and control over using your own infrastructure
Key features summary
- Bulletproof Security: Based on microVMs, provides virtual-machine-level security that fundamentally blocks container vulnerabilities (kernel escapes)
- Instant Startup: Initial boot time under 200ms, achieving code execution startup that is dramatically faster than VMs
- Self-Hosting & Full Control: Can be built and run directly on local or your own servers without cloud lock-in
- OCI compatible: Runs standard container images as-is, so it is compatible with existing Docker and container workflows
- AI-Ready (MCP support) : Integrates and extends naturally with MCP-based AI such as Claude and Agno
Fast development and execution workflow
1. Start the server
- Start the microsandbox server and set up a development environment with a simple command
- The server also acts as an MCP server, so AI tools such as Claude can call it directly
2. Install the SDK
- microsandbox SDKs provided for major languages such as Python, JavaScript, and Rust
- Support for many additional languages (SDK extensibility) ensures broad developer and AI integration
3. Run code safely
- Sandbox environments for several languages such as Python, JavaScript, and Rust can be selected and run separately
- Each sandbox is an independent execution environment, ensuring system safety even when running external code
- SDK examples make it easy to implement asynchronous, automated safe code execution processes
Project-based environment management
- Sandboxfiles (configuration files) are created and managed per project, similar to a developer-friendly package manager workflow
- Multiple sandbox environments (e.g. different languages, settings) can be added to a project and managed by version or environment
- When a project sandbox runs, file and installation changes are automatically preserved in a local directory (./menv)
- Option to enable temporary sandboxes (all history and state completely deleted and isolated when the session ends)
System-wide sandbox installation
- Frequently used environments or apps can be installed and registered as separate executables
- Enter a sandbox execution environment directly from the terminal with a one-line command, without a project path
- Each sandbox can be named and multiple sandboxes with different settings can be run, with session state persisting
Key use cases
AI code execution and development environments
- Quickly provides isolated, reproducible development environments when AI automates actual source code builds, execution, and debugging
- Suited to code automation such as web app generation, bug fixing, and prototyping
Data analysis
- Safely use major data science libraries such as NumPy, Pandas, and TensorFlow inside the sandbox
- Ideal for analysis workflows that need protection, such as those involving personal information or sensitive data
Web browsing agents
- AI safely performs automated tasks such as website navigation, form submission, login, and data crawling
- Useful for content collection, price comparison, automated testing, and more
Instant app hosting
- Tools, demos, calculators, visualizations, and other things users make can be shared instantly as services
- Each app runs in its own isolated space, with support for quickly creating and tearing down temporary environments
System architecture
- Users call the microsandbox SDK from their own business logic
- Requests to pass and run untrusted code go to the server process (microsandbox server)
- Inside the server, each execution request runs in a separate microVM, isolated from the others
- Each microVM can be set up with an independent Python/Node environment
Open-source policy
- Distributed as open source under the Apache License 2.0