OpenSandbox is a secure, fast and extensible sandbox runtime built for AI agents, originally from Alibaba’s open-source ecosystem. The project has more than 16,000 stars on GitHub, is listed in the CNCF Landscape, and was under active development in 2026.
Agents that write and execute code need somewhere safe to run it. OpenSandbox provides isolated environments where agent-generated code and tool calls can execute without touching your host systems or production data. It runs on Docker for local development and on Kubernetes for production clusters, exposes a unified API and SDKs for managing the full sandbox lifecycle, and follows OpenSSF best practices. Teams use it as the execution layer underneath coding agents, data-analysis assistants and other tools that run untrusted, model-written code.
Performance is a headline feature: on Kubernetes, its fast sandbox runtime creates Firecracker-based sandboxes with a median creation time measured in tens of milliseconds, and sandboxes support pause and resume with memory and disk state preserved โ useful when an agent waits on a human or a slow external service. The project targets the scenarios agent platforms actually hit: AI code execution, coding agents, GUI and browser agents, agent evaluation and reinforcement-learning training environments. Multiple language SDKs and a documented architecture make it straightforward to embed as the execution layer of a larger platform.
OpenSandbox is free and open source under the Apache-2.0 license. You run it on your own infrastructure, so costs are limited to the compute your sandboxes consume.
Agent platform builders, infrastructure teams, and vendors of AI coding or data tools that need a safe place to execute model-generated code.