CubeSandbox boots an agent sandbox in under 60ms

Every agent that runs code needs somewhere to run it that is not your machine. The hosted answer is E2B. CubeSandbox is Tencent Cloud's answer, it speaks E2B's SDK, and you run it yourself.

What it is

A sandbox service built on RustVMM and KVM: hardware-isolated microVMs, single node or scaled to a cluster, created in under 60ms with less than 5MB of memory overhead each. Because it is E2B SDK compatible, code written against the hosted incumbent can point at your own cluster instead.

Why it showed up now

12,171 stars, 1,045 in the discovery window, created April 10, pushed September 12, with v0.7.1 tagged the day before. The release notes read as a product roadmap being executed in public: snapshots and copy-on-write clone in 0.3, credential vault and dashboard in 0.4, AutoPause and Terraform deploy in 0.5, Kubernetes deploy and a pluggable volume framework in 0.6.

How it actually works

The 0.7 headline is cross-node pause and resume. With an S3 backend, a sandbox suspended on one node resumes on another, and new sandboxes can be created from snapshots. It is marked preview, and it lets a sandbox fleet be scheduled across nodes instead of pinned to one.

Underneath that sits CubeCoW, a copy-on-write snapshot engine introduced in 0.3, which is what makes event-level snapshots, instant clones and rollback to any saved state affordable at hundred-millisecond granularity.

AutoPause suspends idle sandboxes and wakes them on the next request, so you pay for agent sessions rather than agent wall-clock. The credential vault keeps API keys outside the sandbox entirely: the agent calls external services as usual, and the keys never enter the box it could leak them from. Given prompt injection, that is the most useful item on the list.

0.7 also splits the control plane from operations, moving node management into CubeOps with multi-replica deployment and its own CLI.

QM gives each employee a sandbox inside Slack. CubeSandbox is the layer such a product sits on if you do not want to rent it.

Try it

Deployment is Kubernetes or Terraform, with ARM64 supported natively since 0.5. The quick start lives at cubesandbox.com, and the E2B-compatible SDK means existing client code mostly does not change.

Where it is weak

GitHub's licence detector returns NOASSERTION, which means no standard licence file it recognises. For infrastructure you intend to run in production, resolve that first: read the actual terms in the repository before you plan a deployment around it.

133 open issues, and the README is largely a wall of release announcements rather than an architecture document. Every performance figure quoted here is the vendor's own, measured on unstated hardware.

KVM means bare-metal Linux or nested virtualisation, so this is not something you sprinkle onto an existing managed Kubernetes cluster without checking what the nodes actually are. Cross-node resume needs an S3 backend and is labelled preview.

The open-source project and a hosted product share a domain and a roadmap. The risk is that features you depend on land only in the hosted product.