opensandbox-group/OpenSandbox

Secure, Fast, and Extensible Sandbox runtime for AI agents.

What it solves

OpenSandbox provides a secure, isolated environment for AI applications to execute code, run commands, and interact with files without risking the host system. It solves the challenge of providing AI agents (like coding or GUI agents) with a safe "sandbox" where they can perform tasks, be evaluated, or undergo reinforcement learning training in a controlled manner.

How it works

The platform uses a unified API and multi-language SDKs to manage the lifecycle of sandboxes. It supports multiple runtimes, including Docker for local development and Kubernetes for large-scale distributed scheduling. For enhanced security, it integrates with strong isolation runtimes like gVisor, Kata Containers, and Firecracker microVMs. It also includes an ingress gateway for traffic routing and a credential vault to inject secrets into workloads without exposing them to the the AI agent.

Who it’s for

  • AI Agent Developers: Those building coding agents, GUI agents, or browser automation tools.
  • ML Researchers: People performing agent evaluation or RL training that requires isolated execution environments.
  • Software Engineers: Developers needing a scalable code interpreter or remote development environment (e.g., VS Code in a sandbox).

Highlights

  • Multi-Runtime Support: Works with Docker and Kubernetes, supporting both local and distributed scales.
  • Strong Isolation: Compatible with secure container runtimes like gVisor and Firecracker.
  • Broad Integration: Provides an MCP server for compatibility with clients like Claude Code and Cursor.
  • Comprehensive Tooling: Includes multi-language SDKs (Python, Java, JS/TS, C#, Go) and a dedicated CLI (osb).
  • Built-in Environments: Ready-to-use implementations for code interpreters, Chrome/Playwright browser automation, and full VNC desktops.

Related

  • Project
    e2b-dev/E2B

    An open-source infrastructure that provides secure, isolated cloud sandboxes for running AI-generated code and interacting with virtual desktops.

  • Dispatch
    Securing AI Agents with Isolated Docker Sandboxes

    An exploration of agent-sandbox, a tool designed to run AI coding agents within isolated Docker containers to prevent unauthorized system access and ensure execution safety.

  • Project
    madarco/agentbox

    A tool for running multiple AI agents in parallel within sandboxed VMs or Docker containers, either locally or in the cloud, with persistent IDE and shell access.

  • Dispatch
    Sandboxing AI Agents and Developer CLIs: The Critical Need for Isolation

    An exploration of the disappeared urgency for secure execution environments for AI agents and developer tools, highlighting the challenges of preventing unauthorized system access.

  • Project
    h5i-dev/h5i

    A lightweight, pure-Rust browser for AI agents that provides policy-controlled network access, detailed auditing, and multiple tiers of sandboxing.