CubeSandbox is a high-performance, out-of-the-box secure sandbox service open-sourced by Tencent Cloud, designed specifically for code execution and tool invocation scenarios of AI agents. Built on RustVMM and KVM, the project balances security and performance, supports single-node deployment, can also be conveniently scaled to a multi-node cluster, and is compatible with the E2B SDK.
Purpose CubeSandbox targets code execution and tool execution scenarios for AI agents, providing them with hardware-isolated runtime environments. It can be used to create isolated sandboxes to execute tasks such as code interpreters, helping to strike a balance between security and performance.
Main capabilities According to official materials, CubeSandbox has the following capabilities:
- Millisecond-level startup: cold start on bare metal is 60 milliseconds at single concurrency; under 50 concurrent creations, the average is 67 milliseconds, P95 is 90 milliseconds, and P99 is 137 milliseconds, always staying below 150 milliseconds.
- Low memory overhead: memory overhead is measured when the sandbox specification does not exceed 32GB, and is below 5MB; larger configurations may show marginal growth.
- Hardware isolation: can create hardware-isolated, fully usable sandboxes.
- Compatible with the E2B SDK.
- Flexible deployment: supports single-node deployment and can be easily scaled to a multi-node cluster.
- Web console (WebUI, port :12088): supports visual management from overview to sandbox creation to streaming logs.
- Template capabilities: provides a Template Store, where official preset templates can be installed, and existing READY templates can also be managed.
- Secure proxy and credential vault: includes CubeEgress domain filtering, injection, and auditing.
- Digital assistant AgentHub: can create and manage OpenClaw assistants (Preview).
How to use The official quick start guide includes four steps: prepare the server, install Cube Sandbox, create a sandbox template, and run the first agent code. No build from source is required, and it can run within minutes. After installation, three steps are recommended: check the overview to confirm the node is Ready and capacity is healthy; install official presets from the Template Store (skip this if a READY template already exists); go to Sandboxes → + New sandbox, select a READY template to create a sandbox, and view real-time logs on the details page within seconds. If you do not have KVM access, you can try running Cube Sandbox in a disposable OpenCloudOS 9 virtual machine (QEMU VM). In additiona PVM deployment method is also provided, which can be deployed on ordinary cloud virtual machines without the need for bare metal or nested virtualization.
Applicable scenarios CubeSandbox is suitable for scenarios requiring secure isolation such as code execution and tool execution for AI agents. The official example projects cover code execution, browser automation, OpenClaw integration, RL training, and more. The community also has the Cube 100 program, aimed at the first 100 teams running AI agents in production environments.
Limitations and requirements CubeSandbox requires an x86_64 Linux environment and supports KVM. Memory overhead data is measured when the sandbox specification does not exceed 32GB; larger configurations may show marginal growth. The AgentHub digital assistant feature is marked as Preview. The project is released under the Apache License 2.0 and has been listed in the CNCF Landscape.
