Find your next open-source project

Search by name, purpose, or keyword — or paste a GitHub repository URL

TencentCloud/CubeSandbox

AI Engineering

A hardware-isolated MicroVM sandbox service for AI agents, with sub-60ms cold starts and E2B SDK compatibility.

13,355
1,235
Go
NOASSERTION
7 months old

What it does

CubeSandbox is a sandbox service for running AI agent workloads. Each sandbox is a MicroVM with its own OS kernel, running on KVM and built on RustVMM. The project targets agent platforms that need to execute untrusted code, automate browser sessions, or run training-style tasks without sharing a kernel with the host.

The main pieces are a REST API gateway (CubeAPI), a cluster orchestrator (CubeMaster), a per-node scheduler (Cubelet), an eBPF-based virtual switch (CubeVS), and an OpenResty egress gateway (CubeEgress). Admins can manage nodes, templates, and sandboxes from a web console on port 12088 after installation. For callers, the API is E2B-compatible, so an existing E2B client can point at CubeSandbox by changing a URL setting rather than rewriting client code.

Quick start

Prerequisites are an x86_64 Linux host with KVM available. The README recommends a PVM cloud VM for normal cloud hosts and bare metal for maximum performance. The project warns that the QEMU-based dev environment is not recommended for production because of poor performance.

  1. Deploy on a PVM cloud VM or a bare-metal server following PVM deployment or bare-metal deployment. No source build is required.
  2. Open the web console at http://<control-node IP>:12088.
  3. In Overview, confirm nodes are Ready and capacity looks healthy.
  4. Install a preset from Template Store, or use an existing READY template. Templates are images from which sandboxes are created; see the template guide.
  5. Go to Sandboxes → + New sandbox, pick a READY template, and open the detail page to watch logs within seconds.

The WebUI console guide documents the console flow. For code-level examples, the example projects cover code execution, browser automation, and OpenClaw integration. A Python client package is published as cubesandbox on PyPI.

Deployment and limits

CubeSandbox needs KVM, so it does not run inside a container without nested virtualization. The recommended deployment paths are PVM cloud VMs and bare metal; Kubernetes deployment is marked as a preview in v0.5 and v0.6 changelogs. ARM64 is supported natively across build and deployment.

Performance figures from the README apply to bare metal: under 60ms single-concurrency cold start, and under 50 concurrent creations an average of 67ms with P95 90ms and P99 137ms. Memory overhead is stated as under 5MB for sandbox specifications up to 32GB, with larger configurations expected to be somewhat higher. These are project-reported benchmarks, not independent measurements.

The license is Apache 2.0. The project acknowledges building on Cloud Hypervisor, Kata Containers, virtiofsd, and other upstream components with customized changes.

HotRepo's take

CubeSandbox appeared on the HotRepo daily list with a visible mid-2026 run. Stars rose from 6,733 on 2026-07-01 to 9,193 on 2026-07-09, including three consecutive days with more than 500 stars gained (665, 555, and 564 on July 7–9). It later appears in the Go-language daily list on 2026-11-18 at 13,350 stars with +66 that day and a best Go-list position of 16, which suggests the repository kept growing after the initial spike rather than dropping off. Its repeated Show HN submissions and CNCF Landscape listing fit a project positioning itself as E2B-compatible infrastructure. No GitHub overall daily rank has been recorded, so the evidence supports sustained interest but not a specific peak position.

Written by AI from the project's README, updated . The project's own docs are the source of truth.

Hacker News

Show HN: We built a <60ms, open-source alternative to E2B using RustVMM and KVM
▲ 7 · 2 comments · April 22, 2026

You might also like

  1. pydantic/montyRuns Python code written by an LLM in a sandbox that has no filesystem, network, or environment variables, and starts in under a millisecond.Rust8,769
  2. opensandbox-group/OpenSandboxRun AI agents in isolated Docker or Kubernetes sandboxes to execute code, commands, files, browsers and desktops through one API.Python16,743
  3. kubernetes-sigs/agent-sandboxA Kubernetes CRD and controller for running stateful, singleton sandboxes with stable identity, aimed at AI agent runtimes and RL workloads.Go4,429
  4. devvmhq/devvmPer-project VMs with auto-detected toolchains, ready for human or AI coding agents.Go105