Tencent Cloud open-source CubeSandbox, compatible with E2B, runs 2,000 sandboxes on a single machine

ME News Report, April 21 (UTC+8), according to Beating Monitoring, Tencent Cloud has open-sourced the AI Agent Sandbox Cube Sandbox, written in Rust, under the Apache 2.0 license. The sandbox is an isolated environment for running agent code, preventing models from accidentally deleting files or accessing the host system with excessive privileges. Currently, products like OpenAI Agents SDK, Manus, Perplexity, and Hugging Face are using similar architectures, with the interface standard being E2B. Cube is compatible with the E2B interface; business code doesn’t need to change—by simply modifying an environment variable, an agent can switch from Tencent’s E2B service to a self-deployed Cube. Tencent Cloud released two sets of performance data. Cold start latency under single concurrency is less than 60ms; at 50 concurrency, the average is 67ms, P95 is 90ms, and P99 is 137ms. A single instance consumes less than 5MB of resident memory (measured with sandbox specs not exceeding 32GB), and one 96-core server can run over 2,000 sandboxes simultaneously. In similar scenarios, Docker container startup takes about 200ms using shared host kernel; traditional virtual machines start in seconds, with single-instance memory starting at 20MB. Cube’s approach is to allocate a separate Guest OS kernel for each agent, using hardware-level isolation, while reducing startup time to under a hundred milliseconds. Acceleration relies on resource pool preloading, snapshot cloning, and underlying lock optimizations; memory compression is achieved through rewriting in Rust, copy-on-write (CoW) memory reuse, and reflink disk sharing. The project also includes CubeVS, which uses eBPF for network isolation between sandboxes. Large-scale validation provided two cases: Cube originally ran within Tencent Cloud’s Serverless system, handling over 10 billion calls; after migrating Yuanbao AI programming scenarios to Cube, resource consumption was reduced by 95.8%. Among external clients, MiniMax uses Cube for agentic reinforcement learning (RL) training, achieving minute-level scheduling of hundreds of thousands of sandbox instances. The next step is to open-source event-level snapshot rollback, providing sub-100ms state rollback. (Source: BlockBeats)