CreateOS Sandbox is designed for AI agent builders, offering them dedicated, fast, secure, and hardware-isolated sandbox environments. The primary purpose is to provide a safe and efficient space for developing and running AI agents, ensuring that code execution is contained and controlled.
The challenge in AI agent development often lies in managing the execution of potentially untrusted or complex code. Traditional environments can be slow to set up, lack robust security isolation, or make it difficult to control network egress, posing risks to the host system and data. CreateOS Sandbox addresses this by providing a secure, isolated execution environment that can be provisioned and managed with speed and reliability.
A key feature is the in-kernel eBPF egress control, which enforces allowlists for hosts, IPs, or CIDRs. This ensures that code running within the sandbox can only communicate with approved destinations, preventing unauthorized network access even if the code is compromised. This provides a strong security boundary, as the policy is enforced at the kernel level and cannot be bypassed by user-space processes.
The 'Fork' capability allows for the creation of snapshots of running sandboxes, which can then be cloned in milliseconds. This is invaluable for rapid iteration, testing different agent behaviors, or quickly replicating an environment for parallel processing. The speed of forking means minimal downtime and a seamless development workflow.
CreateOS Sandbox also offers encrypted peer-to-peer mesh networking and VPN capabilities to connect back to your local machine, facilitating secure remote access and communication. Additionally, it supports BYO-S3, allowing users to mount any S3 service as a shared disk across sandboxes, enabling efficient data sharing and persistence where needed.
The product integrates seamlessly with existing workflows through its CLI and SDK. The CLI allows for easy project provisioning and sandbox creation, while the SDK provides programmatic access to sandbox functionalities. The platform also supports BYO-Infra, enabling users to run sandboxes on their own hardware for maximum control and customization.
CreateOS Sandbox operates by providing each AI agent with its own isolated environment, featuring a guest kernel. Network egress is strictly controlled in-kernel using eBPF, ensuring that policies are enforced from outside the sandbox and cannot be circumvented. This approach guarantees that code running inside the sandbox adheres to predefined security rules.
The benefits for users include enhanced security through hardware isolation and kernel-level egress control, significant speed improvements with sub-30ms (p90) sandbox spin-up times, and increased development efficiency through features like instant forking and seamless integration with existing tools. It also offers flexibility with BYO-Infra and BYO-S3 options.
Concrete use cases include running untrusted code for Claude-managed agents, setting up multi-node clusters for distributed AI tasks, performing batch inference, and handling media transcoding like ffmpeg. It can also serve as a self-hosted alternative for GitHub Actions runners, reducing costs and improving CI/CD pipeline performance.
CreateOS Sandbox is targeted at AI agent builders and developers. It offers a CLI and SDK, with over 50 real-world examples available. The product also features a Sandbox plugin for Claude, allowing it to run untrusted code in a self-destructing box when idle. Users can start with 500 free alpha credits without requiring a credit card.
In summary, CreateOS Sandbox provides a secure, fast, and isolated execution environment for AI agents, empowering developers with advanced features for building, testing, and deploying AI applications with confidence and efficiency.