@xfgong: Despite the abundance of open-source Sandbox services, none fully meet my needs. The two core requirements are: 1. The rootfs should not require snapshots, which means it should be cloud storage like virtual machines rather than local disks. This mainly addresses availability un…

X AI KOLs Timeline News

Summary

The author discusses the shortcomings of open-source Sandbox services and proposes two core requirements: rootfs should use cloud storage to improve availability during node failures and reduce pause latency, and the /workspace file interface should be usable even when the sandbox is not started. They also emphasize building on the Kubernetes ecosystem rather than maintaining a separate cluster.

Despite the abundance of open-source Sandbox services, none fully meet my needs. The two core requirements are: 1. The rootfs should not require snapshots, meaning it should be cloud storage like virtual machines rather than local disks. This mainly addresses availability during node failures as well as latency when pausing sandboxes. 2. The /workspace file operations interface should be usable even when the sandbox is not started. This can significantly reduce sandbox startup requirements and lower costs. Ideally it should be built on Kubernetes — some customization is fine, but it shouldn't treat Kubernetes merely as a host (similar to OpenSandbox's fast sandbox or how CubeSandbox is used). Kubernetes has better talent and ecosystem. Otherwise you'd need to operate a dedicated Sandbox cluster, which most companies can't handle — that's also the money E2B makes.
Original Article
View Cached Full Text

Cached at: 10/17/26, 02:59 PM

Even though there are so many open-source sandbox services available, none of them fully meet my needs. The two most critical requirements are:

  1. No snapshot needed for rootfs — the rootfs should be cloud storage rather than a local disk, much like a virtual machine. This is mainly to address availability concerns when a node fails, as well as latency issues when pausing a sandbox.

  2. The /workspace file operations API should work even when a sandbox is not running — this significantly reduces the startup requirements for sandboxes, cutting costs.

Ideally, it should be built on top of Kubernetes. Some degree of customization is acceptable, but it should not merely use Kubernetes as a host machine (like OpenSandbox’s fast sandbox or how CubeSandbox is used). The talent pool and ecosystem around Kubernetes are far superior. Otherwise, you’d need to operate and maintain a dedicated sandbox cluster, which most companies can’t handle — and that’s exactly what E2B makes money from.

Similar Articles

@nash_su: Last night I was chatting with @i5ting about how a sandbox is an indispensable part of any agent, and today Tencent open-sourced Cube Sandbox: 1. Blazing-fast boot (<60 ms) thanks to snapshot cloning and warm resource pools 2. Hardware-grade security via KVM micro-VMs + eBPF network isolation, battle-tested on Tencent Cloud 3. Ultra-high density—runs in <5 MB RAM, thousands per node 4. E2B-compatible

X AI KOLs Timeline

Tencent open-sources Cube Sandbox, a KVM-based micro-VM for agents that boots in <60 ms, runs in <5 MB RAM, and is E2B-compatible.

TencentCloud/CubeSandbox

GitHub Trending (daily)

TencentCloud's CubeSandbox is an open-source, high-performance secure sandbox service for AI agents, built on RustVMM and KVM, offering hardware-level isolation, sub-60ms startup, and compatibility with the E2B SDK.

@aiandcloud: Microsoft has also joined the open source sandbox arena with a cross-platform lightweight microVM implementation. It can run on both Windows and Linux, and is likely the underlying open source implementation of Azure Sandbox.

X AI KOLs Timeline

Microsoft has open sourced NVX, an ultra-lightweight cross-platform MicroVM sandbox built on OpenVMM, designed to run untrusted workloads with hardware-enforced isolation. It supports Windows (WHP) and Linux (KVM/MSHV), and is believed to be the underlying implementation of Azure Sandbox.