Show HN: EdgeBox – A local sandbox that gives your LLM agent a full GUI desktop (github.com)

EdgeBox is a new desktop app that brings a full GUI sandbox to local LLM agents, turning them from code-only executors into digital workers that can drive a browser, edit code in VS Code, click and type in windows, and capture screenshots. Built on the open-source E2B Code Interpreter and exposing functionality via the Model Context Protocol (MCP) HTTP API, EdgeBox adds an integrated VNC-powered Ubuntu desktop alongside a bash shell and isolated code execution. Unlike most OSS sandboxes that are CLI-only, EdgeBox provides both GUI + CLI, multi-language code execution (Python, TypeScript/Node, R, Java, Bash), programmatic mouse/keyboard controls, window management, filesystem ops, screenshots, and secure Docker-backed containers per session. Technically, EdgeBox runs as an Electron frontend (React/TypeScript) with a Node.js backend using Dockerode, requires Docker Desktop, and authenticates tool access through MCP so any MCP-compatible client (Claude Desktop, OpenWebUI, LobeChat, etc.) can connect. Sessions are isolated in separate containers with configurable CPU/memory and network controls and support concurrent x-session-id headers for parallel workflows. For AI/ML practitioners this means low-latency, private, reproducible agent development and realistic “computer use” testing locally—enabling richer automation, web interaction, and multimodal agent research—while also raising practical security considerations around giving agents GUI control (mitigated here by container and network isolation).