Tencent Open-Sources Cube Sandbox

Tencent has open-sourced Cube Sandbox, a runtime for artificial intelligence agents that the company says can start a hardware-isolated session in under 60 milliseconds. (github.com) Cube Sandbox is built on RustVMM and KVM, two virtualization layers used to run lightweight virtual machines, and Tencent says each sandbox adds less than 5 megabytes of memory overhead. The public GitHub repository was active this week, with commits on April 23, 2026 and about 2,500 stars at the time of writing. (github.com 1) (github.com 2) A sandbox is the locked room where an agent runs code, opens files, and reaches the internet without touching the host machine. E2B sells that kind of managed environment for coding agents, data analysis, and other artificial intelligence workloads through its Sandbox software development kit. (e2b.dev 1) (e2b.dev 2) Tencent says Cube Sandbox is compatible with the E2B SDK, which means developers can point existing E2B-based applications at Cube by changing configuration rather than rewriting client code. Cube’s documentation says the switch can be made by changing one environment variable. (github.com) (docs.cubesandbox.ai) That puts Tencent into a part of the agent stack that has become more important as developers give models tools that can execute shell commands, install packages, and browse the web. E2B’s own documentation for the OpenAI Agents SDK describes sandboxed environments as the place where those agents get filesystem, terminal, and network access. (e2b.dev) Tencent’s pitch is density as much as security. The project README says Copy-on-Write memory sharing and a trimmed runtime let one machine run thousands of agents, while Tencent’s Chinese-language developer post says a single server can support more than 2,000 instances. (github.com) (cloud.tencent.com) The company also says the system can scale from one node to a cluster, with checkpoint, restore, and fork features for reusing sandbox state. Its architecture documents describe a runtime made of a shim, hypervisor, and agent beneath a node-level control layer called Cubelet. (github.com) (docs.cubesandbox.ai) Tencent is not alone in trying to standardize the runtime layer for agents, and the company’s performance claims are from its own project materials rather than an independent benchmark. But by publishing the code and matching an existing SDK interface, Tencent is giving developers another way to run agent workloads without depending on a single hosted sandbox provider. (github.com) (e2b.dev) For developers building agents that write and run code, the contest is moving below the model and into the execution layer. Tencent’s bet is that faster startup, tighter isolation, and open-source code will be enough to win a place there. (github.com)