Gru Sandbox

gbox

gbox is a self-hostable sandbox for AI Agents to execute commands, surf web and use desktop/mobile. See “Features” section for details.

This project is based on the technology behind gru.ai. It has been tested over 100000 Agent jobs.

As MCP is getting more and more popular, we also implemented a MCP server to make it easy to be directly integrated into MCP client such as Claude Desktop/Cursor.

Features

1. Terminal
   • Execute any linux command
   • Execute python scripts directly
   • Share session across invokes [under-development]
2. File
   • Mount host machine folders into sandbox
   • Access sandbox files through http links
   • Read file content in multi-modal
   • Write/re-write files
   • Edit files [under-development]
   • Search files [under-development]
3. Browser
   • Open any url, return content in multi-modal
   • Download from any url [under-development]
   • Operate browser by instructions
   • Human take over [under-development]
4. Computer-Using Agent for Android
   • Natural language task execution via ADB client
   • Physical Android devices supported only
5. HTTP Server
   • Start http service on any folder on demand [under-development]
6. SDKs
   • Python SDK: Install using pip install pygbox. See PyPI for details.
   • Typescript SDK
7. MCP
   • Standard MCP support
   • Integrate Claude Desktop & Cursor

Use gbox as a SDK

Python SDK

pip install pygbox

Python SDK Documentation{:target=“_blank”}

Typescript SDK

npm install @gru/gbox

Typescript SDK Documentation{:target=“_blank”}

Use gbox as a CLI

Installation

System Requirements

• macOS 10.15 or later
• Docker Desktop for Mac
• Homebrew

Note: Support for other platforms (Linux, Windows) is coming soon.

Installation Steps

## Install via Homebrew
brew tap babelcloud/gru && brew install gbox

## Initialize environment
gbox setup

## Export MCP config and merge into Claude Desktop
gbox mcp export --merge-to claude
## or gbox mcp export --merge-to cursor

## Restart Claude Desktop

Update Steps

## Update gbox to the latest version
brew update && brew upgrade gbox

## Update the environment
gbox setup

## Export and merge latest MCP config into Claude Desktop
gbox mcp export --merge-to claude
## or gbox mcp export --merge-to cursor

## Restart Claude Desktop

Command Line Usage

The project provides a command-line tool gbox for managing sandbox containers:

## Cluster management
gbox cluster setup    # Setup cluster environment
gbox cluster cleanup  # Cleanup cluster environment

## Container management
gbox box create --image python:3.9 --env "DEBUG=true" -w /app -v /host/path:/app   # Create container
gbox box list                                                                      # List containers
gbox box start <box-id>                                                            # Start container
gbox box stop <box-id>                                                             # Stop container
gbox box delete <box-id>                                                           # Delete container
gbox box exec <box-id> -- python -c "print('Hello')"                               # Execute command
gbox box inspect <box-id>                                                          # Inspect container

## To use the Computer-Using Agent for Android, an OPENAI_API_KEY is required.
gbox cua android "Open Uber and order a ride to The Chinese University of Hong Kong."

## MCP configuration
gbox mcp export                          # Export MCP configuration
gbox mcp export --merge-to claude        # Export and merge into Claude Desktop config
gbox mcp export --dry-run                # Preview merge result without applying changes

Volume Mounts

The gbox box create command supports Docker-compatible volume mounts using the -v or --volume flag. This allows you to share files and directories between your host system and the sandbox containers.

The volume mount syntax follows this format:

-v /host/path:/container/path[:ro][:propagation]

Where:

• /host/path: Path to a file or directory on your host system
• /container/path: Path where the file or directory will be mounted in the container
• ro (optional): Makes the mount read-only
• propagation (optional): Sets the mount propagation mode (private, rprivate, shared, rshared, slave, rslave)

Examples:

## Basic bind mount
gbox box create -v /data:/data --image python:3.9

## Read-only bind mount
gbox box create -v /data:/data:ro

## Multiple bind mounts
gbox box create \
  -v /config:/etc/myapp \
  -v /data:/var/lib/myapp:ro \
  -v /logs:/var/log/myapp:ro:rprivate \
  --image python:3.9

Note: The host path must exist before creating the container. The container path will be created automatically if it doesn’t exist.

Computer-Using Agent on Android

📋 Prerequisites

1. An Android device connected via USB or ADB over TCP/IP
2. ADB (Android Debug Bridge) installed and configured
3. API key for OpenAI with access to the computer-use-preview model

🔧 Setting up ADB

ADB (Android Debug Bridge) is required for gbox to communicate with your Android device:

1. Install ADB:

   • Windows: Download Android SDK Platform Tools and extract the ZIP file
   • macOS: brew install android-platform-tools
   • Linux: sudo apt install adb (Ubuntu/Debian) or sudo pacman -S android-tools (Arch)

2. Add ADB to your PATH:

   • Windows: Add the path to the extracted platform-tools folder to your system’s PATH environment variable
   • macOS/Linux: Add the following to your ~/.bashrc or ~/.zshrc:

     export PATH=$PATH:/path/to/platform-tools
     

3. Verify ADB installation:

   adb version
   

4. Enable USB debugging on your Android device:

   • Go to Settings → About phone
   • Tap Build number 7 times to enable Developer options
   • Go to Settings → System → Developer options (location may vary by device)
   • Enable USB debugging

5. Connect Android device via USB and confirm ADB connected

   adb devices -l
   List of devices attached
   897X0691U       Pixel_3 device
   

6. Set the OPENAI_API_KEY environment variable and enjoy gbox Android CUA

   export OPENAI_API_KEY=YOUR_KEY
   gbox cua android "Open Uber and book a ride to The Chinese University of Hong Kong now."
   

   The video demonstration for this task can be found in the first use case below under “Use Cases” section.

Use Cases

1. Call an Uber

Prompt: Open Uber and book a ride to The Chinese University of Hong Kong now.

2. Use DoorDash/Foodpanda to order food delivery.

MCP Use Cases

Your AI client such as Claude Desktop can use gbox MCP to deliver better results, such as

1. Generating Diagrams

Generate diagrams of Tesla stock prices: https://claude.ai/share/34de8ca3-4e04-441b-9e79-5875fa9fc97a

2. Generating PDFs

Generate PDF of latest AI news: https://claude.ai/share/84600933-dcf2-44be-a2fd-7f49540db57a

3. Analyzing and Calculation

Analyze and compare Nvidia/Tesla market cap: https://claude.ai/share/70c335b7-9fff-4ee7-8459-e6b7462d8994

4. Processing Local Files

Find images in download folder and compress into zip. https://claude.ai/share/f8c4c617-9b32-4062-a8e2-2ab33ef46f42

5. Execute Arbitrary Tasks

Download youtube video: https://claude.ai/share/c2ab6bcb-7032-489f-87d5-cc38f72c2ca9

Develop gbox

Prerequisites

• Go 1.21 or later
• Docker Desktop
• Make
• pnpm (via corepack)
• Node.js 16.13 or later

Build

## Build all components
make build

## Create distribution package
make dist

Running Services

## API Server
make -C packages/api-server dev

## MCP Server
cd packages/mcp-server && pnpm dev

## MCP Inspector
cd packages/mcp-server && pnpm inspect

Contributing

We welcome contributions! Please feel free to submit a Pull Request. For major changes, please open an issue first to discuss what you would like to change.

1. Fork the repository
2. Create your feature branch (git checkout -b username/feature-name)
3. Commit your changes (git commit -m 'Add some feature')
4. Push to the branch (git push origin username/feature-name)
5. Open a Pull Request

Things to Know about Dev and Debug Locally

How to run gbox in dev env instead of the system installed one

1. Stop the installed gbox by gbox cleanup. It will stop the api server so that you can run the api server in dev env.
2. Execute make api-dev in project root.
3. Execute ./gbox box list, this is the command run from your dev env.

How to connect MCP client such as Claude Desktop to the MCP server in dev env

1. Execute make mcp-dev in project root.
2. Execute ./gbox mcp export --merge-to claude

How to open MCP inspect

1. Execute make mcp-inspect in project root.
2. Click the link returned in terminal.

How to build and use image in dev env

1. Execute make build-image-python in project root to build Python image, or make build-images to build all images.
2. Change the image name as needed (e.g., make build-image-typescript for TypeScript image).
3. You may need to delete current sandboxes to make the new image effective ./gbox box delete --all

Why MCP client still get the old MCP content?

1. After you change MCP configuration such as tool definitions, you need to run make build to update the dist/index.js file.
2. You may also need to execute ./gbox mcp export --merge-to claude

License

This project is licensed under the Apache License 2.0 - see the LICENSE{:target=“_blank”} file for details.