Tencent Cloud has fully open-sourced Cube Sandbox under Apache 2.0, releasing the complete production-grade sandbox-as-a-service stack rather than only an SDK. The platform is positioned as a foundational runtime for Artificial Intelligence agent deployment, with native support for the OpenAI Python SDK and E2B SDK. Developers can redirect runtime environments and migrate without code changes, giving teams a path from experimentation to large-scale production use.
Cube Sandbox is built at the hardware virtualization layer and focuses on performance, security, and stability. Cold start is as low as 60ms in real-world scenarios, compared with an industry average of 150ms. The system supports minute-level scheduling of tens of thousands of sandboxes, with platform-level burst scheduling exceeding 100K instances. Tencent Cloud also highlights a triple-layer defense architecture with millisecond-level event snapshots and state rollback, intended to provide an undo mechanism for unpredictable agent behavior. That rollback capability will be launched and open-sourced once fully completed.
Tencent Cloud describes five core technical breakthroughs in the platform. Every sandbox runs a dedicated Guest OS kernel via KVM hardware virtualization, avoiding a shared kernel model. Cold start is <60ms (avg. 67ms, P95 = 90ms at 50 concurrent). Per-instance memory overhead <5MB, and CoW sharing, Rust-based trimming, and reflink disk sharing enable 2,000+ sandboxes on a single 96-vCPU host, with 90%+ storage savings versus traditional approaches. Distributed scheduling and bin-packing deliver platform-level burst scheduling of 100K+ instances, with P99 latency below 200ms under 100 concurrent launches on a single 96-vCPU host. Sub-hundred-millisecond snapshots support checkpoint saving, arbitrary state rollback, and rapid forking, though this feature is still being completed for release.
The platform is aimed at three user groups across the agent lifecycle. For foundation model labs, Cube is designed for extreme concurrency in agentic reinforcement learning training, including heterogeneous sandbox environments. For agent developers and small-to-medium businesses, Tencent Cloud says setup can be completed with a one-click script in minutes, with integration available through MCP, API, SDK, or CLI and no code rewrites. For enterprise customers, full private deployment is intended to keep data within enterprise boundaries while meeting cybersecurity and compliance requirements.
Tencent Cloud says Cube Sandbox is already in production and has released its full codebase, deployment scripts, documentation, and examples covering shell execution, file operations, browser automation, and reinforcement learning training. The company plans to pair Cube with TACO Artificial Intelligence Acceleration Engine and FlexKV cache system to create a broader stack combining secure sandboxing, inference acceleration, and cache optimization for large-model applications.
