Tencent CubeSandbox: an E2B-compatible sandbox for AI

What Tencent actually released

What stands out here is not just another sandbox, but an attempt to make the E2B API a practical portability layer. In its open-source CubeSandbox announcement, Tencent AI describes the project as an environment for safely executing AI-agent code with hardware-backed isolation through RustVMM and KVM. This goes beyond an ordinary container boundary: untrusted code is separated through virtualization.

The most practical detail is the claimed compatibility with the E2B SDK. According to Tencent, an existing E2B application can be switched to CubeSandbox by changing a single environment variable, without touching business logic. The project lists a Python SDK, TypeScript SDK, REST API and CLI for access.

At launch, Tencent also claimed cold starts below 60 ms and memory overhead below 5 MB. Those numbers are compelling, especially for agent loops that repeatedly create and destroy sandboxes. Still, they are vendor-reported figures rather than an independent comparison with alternatives.

OpenSandbox sits nearby as another project claiming E2B compatibility. Its emphasis differs: self-hosting, Docker and execution within a private network when code or data cannot leave the organization. Public OpenSandbox materials do not provide comparable cold-start figures, so an honest speed comparison is not yet possible.

Why compatibility matters more than a polished benchmark

The bigger change is not the 60 ms figure, but reduced lock-in between an agent application and a single execution provider. If the E2B SDK becomes a shared layer, teams can choose a sandbox engine based on its isolation model, deployment location and operational constraints instead of rewriting the entire execution path.

For sensitive data, OpenSandbox is an understandable candidate because of its self-hosted model. CubeSandbox is more interesting where hardware isolation and rapid startup matter more than the simplicity of a container stack. Agent developers benefit, while platforms will need to compete on actual runtime reliability rather than API shape.

I would test parallel-load behavior, session-state cleanup and the completeness of E2B compatibility around failures and timeouts before focusing on a demo cold start. That is where a drop-in replacement usually meets engineering reality. If compatibility survives those edge cases, the sandbox market may finally gain a portable interface instead of another closed island.

We previously covered Pydantic Monty, a secure Python interpreter for running LLM-generated code without containers. Its approach provides useful context for evaluating the sandboxing model behind Tencent’s OpenSandbox infrastructure.