OpenSandbox

OpenSandbox is an Agentic AI Foundation (AAIF) open-source project that gives AI applications isolated sandbox environments on your own infrastructure to execute code, run commands, manage files, and operate browsers or desktops. Start locally with Docker and deploy on Kubernetes via a unified sandbox API. Key capabilities include a Firecracker-based Fast Sandbox runtime with high-density, low-latency creation (64ms P50) and pause/resume preserving memory and disk state, ingress/egress network control, a Credential Vault, resource pools and batch creation for evaluation and RL training, plus multi-language SDKs (Python, Java/Kotlin, TypeScript, C#/.NET, Go), an osb CLI, and an MCP server.

• Firecracker-based Fast Sandbox runtime (64ms P50 creation, pause/resume with state)
• Unified sandbox API from local Docker to Kubernetes cluster
• Per-sandbox ingress/egress network control and Credential Vault
• Multi-language SDKs (Python/Java/TS/C#/Go), osb CLI, and MCP integration
• Extensible sandbox protocol with resource pools and batch creation for agent evaluation and RL