I'm the creator of EdgeBox. For a while now, I've been fascinated by building LLM agents, but I've consistently hit a wall: most agent sandboxes are just command-line interpreters. They can run code, but they can't use a computer in the way a human does.
That's why I built *EdgeBox*. It's a free, open-source desktop application that brings the power of a cloud sandbox to your local machine, with one key difference: it provides both a *command-line shell AND a full graphical (GUI) desktop environment*.
The goal is to empower LLMs with true "Computer Use" capabilities. Instead of just executing Python scripts, your agent can now:
* Open Google Chrome to browse websites and research information.
* Use a pre-installed VS Code to interact with a codebase.
* Programmatically control the mouse and keyboard to interact with any desktop application.
* "See" the screen via built-in screenshot capabilities to understand the GUI context.
*How it works:*
The architecture is straightforward. When you start a session, it spins up an isolated Docker container with a complete Ubuntu desktop environment. All communication between your LLM client and the sandbox happens locally via the *MCP (Model Context Protocol)*, a standardized HTTP interface. This means you can easily connect it to clients like Claude Desktop, OpenWebUI, LobeChat, and more.
*Key Features:*
* *100% Local & Private*: All data stays on your machine.
* *GUI + CLI*: The best of both worlds for agent capabilities.
* *Multi-Session Support*: Run multiple, isolated sandboxes concurrently using the `x-session-id` header.
* *Secure*: Each sandbox runs in a separate, resource-limited Docker container.
The project is heavily inspired by the open-source E2B Code Interpreter, which provided a fantastic foundation. My contribution was to package it into a local-first desktop app and, most importantly, add the full GUI desktop layer.
The project is open-source, and I would love to get your feedback, hear your ideas for new features, or see what kind of powerful agents you build with it.
bigppwong•4mo ago
I'm the creator of EdgeBox. For a while now, I've been fascinated by building LLM agents, but I've consistently hit a wall: most agent sandboxes are just command-line interpreters. They can run code, but they can't use a computer in the way a human does.
That's why I built *EdgeBox*. It's a free, open-source desktop application that brings the power of a cloud sandbox to your local machine, with one key difference: it provides both a *command-line shell AND a full graphical (GUI) desktop environment*.
The goal is to empower LLMs with true "Computer Use" capabilities. Instead of just executing Python scripts, your agent can now:
* Open Google Chrome to browse websites and research information. * Use a pre-installed VS Code to interact with a codebase. * Programmatically control the mouse and keyboard to interact with any desktop application. * "See" the screen via built-in screenshot capabilities to understand the GUI context.
*How it works:*
The architecture is straightforward. When you start a session, it spins up an isolated Docker container with a complete Ubuntu desktop environment. All communication between your LLM client and the sandbox happens locally via the *MCP (Model Context Protocol)*, a standardized HTTP interface. This means you can easily connect it to clients like Claude Desktop, OpenWebUI, LobeChat, and more.
*Key Features:*
* *100% Local & Private*: All data stays on your machine. * *GUI + CLI*: The best of both worlds for agent capabilities. * *Multi-Session Support*: Run multiple, isolated sandboxes concurrently using the `x-session-id` header. * *Secure*: Each sandbox runs in a separate, resource-limited Docker container.
The project is heavily inspired by the open-source E2B Code Interpreter, which provided a fantastic foundation. My contribution was to package it into a local-first desktop app and, most importantly, add the full GUI desktop layer.
The project is open-source, and I would love to get your feedback, hear your ideas for new features, or see what kind of powerful agents you build with it.
You can check out the GitHub repo here for downloads and more details: https://github.com/BIGPPWONG/edgebox
Thanks for reading!