Run untrusted AI-generated code safely inside a container that cannot harm the host machine or escape its boundaries.
Give an AI agent a live browser environment to automate web tasks using Chrome and Playwright inside a sandbox.
Deploy many parallel AI agent sandboxes on Kubernetes for a multi-agent workflow with strong security isolation.
Control sandbox network access to restrict which outbound connections an AI agent is allowed to make.
Requires Docker or a Kubernetes cluster, stronger isolation modes like gVisor and Kata Containers need additional host-level setup.
OpenSandbox is a platform from Alibaba for running AI agents inside isolated, controlled environments. When an AI agent needs to execute code, browse the web, manipulate files, or interact with a desktop, it should do so in a contained space where it cannot accidentally harm the host machine or escape its boundaries. OpenSandbox provides that containment layer. The platform supports running sandboxes through Docker on a single machine or through Kubernetes for larger deployments where many sandboxes need to run in parallel. It includes built-in environments for running shell commands, reading and writing files, and executing code through an interpreter. It also ships example configurations for browser automation using Chrome and Playwright, and for graphical desktop environments accessible over VNC or through VS Code. For stronger security isolation, OpenSandbox supports plugging in specialized container runtimes like gVisor, Kata Containers, and Firecracker. These create a deeper barrier between sandbox workloads and the underlying host system, which matters when running untrusted or AI-generated code. SDKs are available in Python, Java, Kotlin, JavaScript, TypeScript, C# and Go, so you can connect to and control sandboxes from whichever language your application uses. There is also a command-line tool called osb for manual sandbox operations: creating sandboxes, running commands inside them, moving files, and managing network access rules. Network control is built in. You can configure what traffic each sandbox is allowed to send or receive, with an ingress gateway for routing and per-sandbox egress controls for restricting outbound connections. The project is listed in the CNCF Landscape, a registry of cloud-native infrastructure tools maintained by the Cloud Native Computing Foundation.
← alibaba on gitmyhub — every repo by this author, as a profile.
Verify against the repo before relying on details.