CreateOS Sandbox

CreateOS Sandbox is designed to provide AI agent builders with dedicated, fast, secure, and hardware-isolated environments. These sandboxes are built to offer a robust platform for developing, testing, and deploying AI agents, ensuring that each agent operates within its own secure boundary. The core problem CreateOS Sandbox addresses is the inherent security risk and complexity associated with running untrusted code, especially within AI agents that may interact with external systems. Traditional methods often lack sufficient isolation or introduce significant overhead, making it challenging to ensure security without compromising performance. This product aims to solve that by providing a secure, isolated execution environment that is both fast and reliable. One of the key features is the in-kernel eBPF egress control, which enforces network policies at a low level, making it extremely difficult for compromised code to bypass security measures. This allows users to define allowlists for hosts, IPs, or CIDRs, ensuring that agents can only communicate with approved destinations. Another significant capability is the 'Fork' functionality, which enables users to snapshot a running sandbox and clone it in milliseconds. This is invaluable for rapid iteration and testing, as it allows for quick duplication of environments with their current state. CreateOS Sandbox also offers encrypted P2P mesh networking and VPN capabilities, providing secure communication channels back to the user's local machine. For data persistence and sharing, the 'BYO-S3' feature allows any S3 service to be mounted as a shared disk across sandboxes, facilitating data access and management. The product integrates seamlessly with existing workflows through its CLI, allowing projects to be provisioned with credentials directly into the .env file using a simple command. Furthermore, the 'BYO-Infra' option empowers users to run these sandboxes on their own hardware, offering flexibility and control over their infrastructure. The platform has been demonstrated with over 51 real-world examples, including Claude-managed agents, multi-node clusters, batch inference, and media transcoding, showcasing its versatility. A recent addition is a Sandbox plugin for Claude, enabling it to run untrusted code in a sandbox that self-destructs when idle, further enhancing security and resource management. CreateOS Sandbox operates by providing each AI agent with its own isolated environment, featuring a dedicated guest kernel. The egress control is enforced in-kernel using eBPF, preventing any code running inside the sandbox from circumventing the defined policies, even if fully compromised. This approach ensures a high level of security and isolation. The benefits for users include enhanced security for running untrusted code, significantly faster development and testing cycles due to rapid forking and snapshotting, and greater control over network egress. The ability to run on BYO-Infra also offers cost savings and compliance advantages. The platform's speed, with p90 cold starts around 30ms, allows for more granular task execution rather than just session-based operations. Concrete use cases for CreateOS Sandbox include deploying AI coding agents that require secure execution environments, running CI/CD pipelines with ephemeral test runners for faster builds and reduced costs, batch processing of data or media, and managing multi-node AI agent clusters. The Claude plugin also enables secure execution of code snippets generated by the AI. CreateOS Sandbox is targeted at AI agent builders and developers. It offers a CLI for integration, a dashboard for management, and SDKs for programmatic access. The product is available with free alpha credits and supports running on user-provided infrastructure. Specific tech stack details beyond eBPF and kernel isolation are not explicitly detailed, nor are formal pricing tiers beyond the initial free credits. In summary, CreateOS Sandbox provides a secure, high-performance, and flexible environment for AI agents, leveraging advanced isolation techniques and developer-friendly tools to streamline the development and deployment process.