CubeSandbox

Alva PoC'd CubeSandbox on GKE. Helm wasn't the hard part. 3 real blockers: – Workload Identity isn't in Docker's keychain – a dead guest snapshot saved as READY – guest DNS dropped by egress policy All fixed upstream. p50 create 228ms cubesandbox.com/blog/posts/2…

Swapping the sandbox backend under a production Agent platform took ~300 lines and one day. Because CubeSandbox is E2B-compatible — the existing backend gets inherited, 4 methods overridden. The real cost wasn't code. It was the pitfalls. cubesandbox.com/blog/posts/2…

🎉Cube Sandbox v0.7.3 is out！ Intermittent network failures are the worst kind of bug — no error, just "sometimes unreachable". This release fixes intermittent network failures when accessing sandboxes, along with the root causes behind them. 🔌 New: sandbox-native MCP support (E2B-compatible). Pass mcp when creating a sandbox — the SDK launches the gateway and hands you the URL and Token. Python / Node / Go all supported. ⚡ S3lvol hot upgrades now pause sandbox I/O for ~8s instead of ~40s. Single pauses are ~41% faster (150ms→88ms for a 2 GiB sandbox). Changelog👉github.com/TencentCloud/Cube… @tencentcloud @TencentAI_News

The thing that most often goes wrong when you give an agent an execution surface isn't the code — it's security. We got OpenClaw + DeepSeek Harness running on a real cluster: • 134ms to spin up an isolated sandbox • Pause/Resume in seconds • A clean split between the control plane (agent sessions / approvals / identity) and the execution plane (isolated sandboxes / resources / short-lived identity) Full benchmark data + a production go-live checklist included. cubesandbox.com/blog/posts/2…

🌍 CubeSandbox is hiring globally! We're building the foundation for AI Agents worldwide — and there's a seat for you. 5 months after open-sourcing: 12,000+ ⭐ on GitHub. Now we're scaling to more countries and regions, looking for new teammates. 📍 Location: Shenzhen, China | Tencent Cloud CSIG 🔧 Role: R&D Engineer, CubeSandbox Elastic Compute Platform What we're looking for: ▫️ BS+ in CS or related field, 3+ years of experience ▫️ Solid systems programming skills (Rust / Go / Python), with a taste for clean, maintainable code ▫️ Deep hands-on experience in at least one of: architecture / networking / storage / virtualization / scheduling — or an obsession with performance Interested? Send your resume to 📩 [email protected] RTs appreciated — help us find the right person 🙌 @tencentcloud @TencentAI_News

Cross-node recovery has a hidden tax. When an agent sandbox gets restored on a new node, its disk data is pulled from object storage on demand — scattered small reads, slow paths back to origin, every time. Cube Sandbox v0.7.2 reworks that read path: immutable objects are now cached locally on the node, and scattered reads are merged into whole-object GETs with a bounded fetch budget. Re-importing the same snapshot later hits the local cache directly. The result: cross-node recovery — or creating from a snapshot — reaches running in under 1 second on a default MinIO setup. Also in this release: a JuiceFS volume plugin with full POSIX semantics, automatic reclaim of idle guest memory so nodes don't stay pinned at peak usage, faster cold starts, hostNetwork by default on K8s, and 25+ fixes. Coming next: a public quality status page — nightly SDK-compat E2E and performance benchmarks on master, results open to everyone. Changelog👉github.com/TencentCloud/Cube… @tencentcloud @TencentAI_News @tenctcloud_Intl

Open-source maintainers know this feeling well: a stranger takes your project seriously enough to measure it properly. Deepak Shah did exactly that — same host, 300 trials, his own production workloads, every output verified. Cube Sandbox came out 5.5–13.3× faster on cold start. And he documented, just as carefully, the cases where Docker wins. We're grateful for both halves of that. To a young project, an honest independent benchmark is worth more than any marketing budget. Thank you, Deepak. If you're evaluating sandbox options, read this one first👉: deepakshah.net/posts/cubesan…

CubeSandbox v0.7.1 is out 🧊 The control plane no longer has single points of failure: template builds moved into a standalone CubeTemplateCenter, CubeMaster runs multi-replica, and the lifecycle manager fails over active-standby. Also fixed one agent builders will feel: slow-first-byte upstreams (LLM traffic) through the transparent proxy no longer get cut off with a 504 after 60s — timeout raised to 2 hours. Full changelog👉: github.com/TencentCloud/Cube… @tencentcloud @TencentAI_News

🚀 CubeSandbox v0.7.0 is here. The biggest ask from production teams: "Why can't I resume my paused sandbox on a different node?" Now you can. Plus: ✅ Components coexist across versions — upgrade without downtime ✅ Faster network provisioning for high-density clusters ✅ Control plane & ops architecture separation Node drain no longer means losing sandbox assets. 🔗 github.com/TencentCloud/Cube… @TencentAI_News @tencentcloud @tenctcloud_Intl

How did Lenovo's AI Lab cut their cloud Agent's sandbox startup from 10s to under 100ms? They migrated from Daytona to CubeSandbox — and got more than speed: self-hosted deployment, an adapter-layer migration strategy, and a sandbox that evolved from disposable tool container to a stateful runtime with snapshot & rollback. Full story: cubesandbox.com/blog/posts/2… @tencentcloud @TencentAI_News @tenctcloud_Intl

New on the Cube blog: how Lexmount put a browser runtime inside an agent sandbox👉cubesandbox.com/blog/posts/2… Lexmount's Insight Flow runs a full browser as an agent execution environment — that means the sandbox needs real internet access, stable startup under batch load, dynamic state injection, and trustworthy capacity signals. Four hard requirements, four real problems hit in production, four real fixes — gateway MAC misresolution on multi-neighbor NICs, startup-order race between network-agent and Cubelet, env vars lost to the snapshot-restore path, and phantom capacity readings from mock metrics. @tencentcloud @TencentAI_News @tenctcloud_Intl

Noticed last week that Cube Sandbox landed in crabbox as one of its providers. crabbox is OpenClaw's remote test execution control plane — warm a box, sync the diff, run the suite — already wired to E2B, Modal, Cloudflare and more. One more backend to switch to when running agent/code tests. PR by @steipete , thanks👋

🎉 Cube Sandbox v0.6.0 is out ☸️ K8s deployment + 💾 Volume framework land — six core features go production-ready ☸️ Control plane + compute nodes via Helm Chart, into any K8s cluster 💾 E2B-compatible Volume framework, bring your own storage backend Template aliases · Configurable inbound Host · Compute node isolation · CubeOps as a dedicated service 🚧 Next: CRD/Operator native mgmt · cross-node pause & resume · full-path HA 👉github.com/TencentCloud/Cube… @TencentGlobal @TencentAI_News @tencentcloud @tenctcloud_Intl @TencentHunyuan

📣 Cube 100 Founding Users Program is live.🚀 We're looking for the first 100 teams putting AI agents into production with Cube Sandbox — and we want you to tell that story. If you're running Cube in production (or getting close), we'd love to co-write your case with you: 1:1 engineering support, official case page on our site, joint promotion, roadmap seat, and a speaker slot at the Global AI Infra OSS Summit. Show your case. Tell your story! ➡️forms.gle/Wjy53fksUoc3y4388 @tencentcloud @TencentAI_News @tenctcloud_Intl

Thank you for helping Cube be seen🍻 @TencentAI_News @tencentcloud

🎉 CubeSandbox v0.5.1 is out — taking production-grade sandboxes one step further. This release ships three core updates: 🧩 Standalone Lifecycle Manager: AutoPause coordination logic is extracted from CubeProxy into a standalone control-plane service, enabling CubeProxy to scale horizontally across multiple replicas. ⏳ Tri-state Timeout Semantics: omit = cluster default, NEVER_TIMEOUT = never reclaim, 0 = reclaim immediately. E2B-aligned. 🔒 Host-Mount Whitelist: mount paths are restricted to a configurable prefix, blocking arbitrary bind-mounts. Plus: a new Node.js/TypeScript SDK (preview) and a fix for the ARM64 64KB-page snapshot issue. Update and give it a try！ Repo: github.com/TencentCloud/Cube… changelog: github.com/TencentCloud/Cube… @TencentAI_News @tencentcloud