EdgeBox: A Local Sandbox That Gives AI Agents Full Computer Access
Most AI agent setups give you one of two things: a code interpreter or a browser automation tool. Rarely both, and almost never running locally.
EdgeBox takes a different approach. It spins up a full Ubuntu desktop environment inside a Docker container on your local machine and exposes everything — code execution, shell access, GUI automation — to LLMs through the MCP (Model Context Protocol) interface.
What You Get
EdgeBox provides two categories of tools:
Code Execution (Always Available)
- Multi-language support: Python, JavaScript, R, Java
- Persistent bash shell with full filesystem access
- File read/write/list/monitor operations
- Isolated inside Docker — no risk to the host system
Desktop Control (When Enabled)
- Full Ubuntu desktop accessible via VNC
- Pre-installed apps: Chrome, VS Code, etc.
- Programmatic mouse control (click, drag, scroll, move)
- Keyboard input and key combinations
- Screenshot capture for visual perception
This means you can instruct an AI agent to:
- Open Chrome, search for something, and save the results to a file
- Run a data analysis script and screenshot the output
- Launch VS Code and edit a specific file
The agent sees the screen, moves the mouse, types on the keyboard — just like a human would.
MCP-Native Integration
EdgeBox speaks MCP over HTTP, which means it works out of the box with any MCP-compatible client:
- Claude Desktop — Anthropic's official desktop client
- OpenWebUI — Self-hosted LLM frontend
- LobeChat — Popular open-source LLM UI
Configuration is straightforward: point your MCP client to EdgeBox's HTTP endpoint. Multi-session support (via session IDs) lets you run multiple sandboxes concurrently without interference.
Why Local Matters
There are cloud-based agent sandboxes out there. EdgeBox's value proposition is running entirely on your machine:
- Privacy: Your code and files never leave your computer. For proprietary codebases or sensitive data, this is non-negotiable.
- Latency: Tool calls resolve over localhost. No network round-trips, no waiting for cloud containers to spin up.
- Isolation: The Docker container keeps agent activity contained. Configurable resource limits prevent runaway processes from affecting your host.
Getting Started
Prerequisites: Docker Desktop installed and running.
Download the installer for your platform from the Releases page:
- Windows:
.exe - macOS:
.app - Linux:
.AppImage/.deb/.rpm
On first launch, EdgeBox pulls the required Docker image automatically.
Who This Is For
- AI agent developers who need a secure execution environment
- Anyone wanting LLMs to perform repetitive desktop tasks (scraping, batch screenshots, automated testing)
- Privacy-conscious users who don't want to send code to cloud sandbox services
- Claude Desktop / OpenWebUI / LobeChat users looking to add "hands-on" capabilities
Summary
EdgeBox fills a clear gap: giving AI agents the ability to actually use a computer, not just talk about it — and doing it locally. MCP protocol compliance, Docker isolation, desktop GUI control, and multi-language code execution make it a complete local workstation for AI agents.
Project: github.com/BIGPPWONG/EdgeBox