Alibaba Launches CoPaw and OpenSandbox: The New Open-Source Workstation for AI Agents
Alibaba’s research team unveiled CoPaw and OpenSandbox, new open-source frameworks designed to scale local AI agent workflows and provide secure, isolated execution environments seamlessly.
Context
The development of autonomous AI agents has grown exponentially, but one of the biggest bottlenecks for developers is the execution infrastructure: how do you run AI-generated code and complex workflows securely, with access to memory and tools, without getting locked into expensive cloud platforms? Until now, creating efficient “sandboxes” or controlled environments was a complex challenge, often resulting in “environment drift” (when an agent works perfectly in dev but fails in production).
Content
Alibaba has addressed this challenge head-on by releasing CoPaw (Co Personal Agent Workstation) and OpenSandbox.
CoPaw acts as a high-performance workstation for deploying and managing personal AI agents. It uniquely combines multi-channel chat integration, local model support, and complete user control into a single open-source hub. This empowers developers to efficiently manage an agent’s memory and scale advanced workflows.
Accompanying CoPaw, the release of OpenSandbox provides the critical technical foundation where these agents operate. According to the official announcement and its GitHub repository, OpenSandbox is a general-purpose execution platform tailored for AI applications. Its key highlights include:
- Robust Environment Isolation: It utilizes Docker for local development and execution, and leverages Kubernetes for distributed, large-scale deployments.
- Broad Ecosystem Integration: It is engineered to plug directly into leading AI frameworks and tools, including LangGraph, Claude Code, Gemini CLI, along with automation libraries like Playwright and Chrome.
- Cost-Free Alternative: Released under the Apache 2.0 license, OpenSandbox eliminates the reliance on managed sandbox services that charge per-minute fees or impose vendor lock-in. It allows software engineers to build, test, and run Coding Agents and GUI Agents directly from their local machines.
These open-source releases by Alibaba’s AgentScope team highlight a clear industry shift toward democratizing not just the model weights (via local LLMs), but the crucial infrastructure needed for AI to act securely in the real world.