agent-infra/sandbox

This repository provides a ready-made, self-contained environment for running AI agents that need to browse the web, execute code, manage files, and interact with development tools. Everything runs inside a single Docker container, which is a standard way to package software so it works the same way on any computer. Starting it takes one command, and within seconds you have a browser, a terminal, a code editor, a notebook environment, and a file system all sharing the same storage space. The key problem this solves is that most sandboxes offer only one capability at a time: you get either a browser or a shell or a code runner, but not all of them together. When an AI agent needs to download a file in the browser and then process it with code, traditional sandboxes make that coordination awkward because each part runs in isolation. Here, all components share one file system, so a file downloaded through the browser is immediately available to a script running in the terminal. Once the container is running, you can open a visual browser through your own web browser using a remote desktop view, write and run code in a full VSCode editor, execute Python notebooks, and use a terminal, all from a standard browser tab on your computer. Developers can also control the sandbox programmatically using Python, TypeScript, or Go libraries that the project provides. AI agents connect to the sandbox through a protocol called MCP (Model Context Protocol), which gives them a standardized way to issue instructions like "go to this URL", "read this file", or "run this command". The container comes pre-configured with those connections, so no manual wiring is required. This project is aimed at developers building AI agents or automated workflows that need a safe, isolated place to do real work on the web and with files, without risking the host machine.