CubeSandbox

Overview

CubeSandbox is Tencent Cloud’s open-source answer to the agent sandbox problem: how do you let untrusted AI-generated code run safely at scale without sacrificing startup latency or burning through resources? The platform creates hardware-isolated microVMs using RustVMM and KVM that boot in under 60 milliseconds and consume less than 5MB of memory each. That performance profile means you can run thousands of concurrent sandboxes on a single node — a requirement for multi-agent workflows where each agent needs its own isolated execution environment.

Key Capabilities

Performance: Sub-60ms cold start times put CubeSandbox in the same class as Firecracker-style microVMs but with an agent-specific feature set. The minimal memory footprint (under 5MB per instance) means a single 64GB node can comfortably run thousands of concurrent sandboxes without paging.

Hardware isolation: Each sandbox gets its own guest OS kernel, not just a container namespace. This provides a stronger security boundary than Docker-based isolation — a compromised sandbox cannot escape to the host through kernel exploits that affect container runtimes.

E2B compatibility: CubeSandbox implements the E2B SDK interface, so teams already using E2B’s sandbox API can migrate without rewriting their agent orchestration code. This is a deliberate design choice — E2B has become a de facto standard for agent sandbox APIs.

Security controls: A credential vault injects secrets into sandboxes at runtime without exposing raw API keys to the guest environment. Domain allowlists and egress traffic auditing let operators control what external services a sandbox can reach, preventing data exfiltration from compromised agent code.

Operational features: Snapshot and rollback at millisecond granularity enables branching execution — take a snapshot before a risky operation and roll back if it fails. The template system pre-installs common dependencies (Python, Node.js, system tools) and auto-distributes templates across cluster nodes for consistent environments.

Use Cases

Agent orchestration platforms use CubeSandbox as the execution layer where AI-generated code runs safely. Teams building coding agents need isolated environments for each task — CubeSandbox provides that at a fraction of the resource cost of full VMs or heavyweight containers. CI/CD pipelines use it for running AI-generated tests and builds in isolation. Research teams working with untrusted model outputs use the hardware isolation to prevent sandbox escapes.

Considerations

CubeSandbox requires a Linux host with KVM support — it will not run on macOS, Windows, or cloud instances without nested virtualization enabled. The project is relatively new and backed by Tencent Cloud, so community support is still growing. The E2B compatibility layer covers the core API but may not support all E2B extensions. Multi-node cluster deployment adds operational complexity that may not be justified for small-scale agent workloads.

Who It’s For

CubeSandbox is built for platform teams running multi-agent systems at scale who need hardware-isolated sandboxes with sub-100ms startup times. It suits organizations that need stronger isolation guarantees than Docker provides, teams migrating from E2B who want to self-host their sandbox infrastructure, and any deployment where thousands of concurrent agent execution environments are required on limited hardware.