OpenSandbox
A general-purpose, secure sandbox platform for AI applications, giving coding agents, GUI agents and code-execution workloads an isolated place to run with the same API from a laptop to a large cluster.
OpenSandbox is a general-purpose sandbox platform for AI applications, providing a secure, scalable place to run coding agents, GUI agents, agent evaluation, AI code execution, and reinforcement learning training workloads.
It exposes one API surface for creating sandboxes, running commands and handling files, through native SDKs for Python, Java, Kotlin, TypeScript, C#/.NET and Go, a CLI, and an MCP server, with the sandbox lifecycle and execution defined as public OpenAPI contracts. Docker and Kubernetes runtimes sit behind the same calls, with built-in command, filesystem and code interpreter environments covering coding agents, browser automation and desktop environments such as VNC and VS Code.
Deployments can mix long-running Kubernetes workloads with short-lived, Firecracker-backed microVM sandboxes for near-constant-time startup, and include a unified ingress gateway, per-sandbox egress controls and a credential vault that keeps real secrets away from sandbox workloads. It is written in Python and released under the Apache 2.0 license.