Mcp Sandbox

The MCP Sandbox is a lightweight, Docker‑based execution environment that lets AI assistants and developers run arbitrary Python code safely and reproducibly. By exposing a set of MCP tools, the server solves the common problem of executing untrusted code without risking host system integrity or resource contention. Each sandbox is a self‑contained container that can be created, reused, and destroyed on demand, ensuring isolation while still allowing persistent state across tool invocations.

At its core, the server offers a rich toolbox for Python developers working with LLMs. The tool spins up a new container and returns an identifier that can be reused for subsequent actions. runs code snippets inside the container, capturing stdout and stderr for immediate feedback. Package management is handled through and , allowing dynamic installation of third‑party libraries on the fly. Terminal commands can be executed via , and files generated by code are automatically exposed through web links thanks to the file generation feature. The server also provides a simple file upload API () to seed containers with local data.

Developers can integrate the sandbox into AI workflows in several ways. For example, an LLM can be prompted to write a data‑analysis script, install necessary libraries, and then execute the script—all within a single conversational turn. The sandbox’s SSE endpoint makes it trivial to stream results back to the assistant, enabling real‑time debugging or iterative refinement. Because each sandbox is a Docker container, it can be scaled horizontally or run on cloud platforms with minimal configuration changes.

Unique advantages of the MCP Sandbox include its zero‑trust execution model—code never runs on the host machine—and its package isolation, which prevents dependency clashes across different user sessions. The integration with the popular Viby tool further extends its capabilities, allowing higher‑level orchestration of complex workflows. Whether you’re building a code‑generation API, a data‑science playground, or an educational platform, the MCP Sandbox provides a secure, repeatable, and developer‑friendly foundation for running Python code in conjunction with AI assistants.