glovebox-mcp — Claude Code MCP server: setup
Runs browser and desktop control inside a nested X11 window for AI agents. Use it when automation must stay away from your real screen and apps.
Open on GitHubWhat glovebox-mcp does
glovebox-mcp exposes MCP tools for a browser or desktop app running inside a Xephyr window on display :1. An agent can inspect screenshots, locate controls, click, type Unicode text, scroll, open local files, and upload browser files through the Chrome DevTools Protocol. Multiple sandbox instances can be managed as well.
The installer offers none, basic, and local vision modes. basic uses Tesseract OCR and returns text with coordinates; local uses OmniParser and requires an NVIDIA GPU. The setup expects Linux, Python 3.10+, uv, Chromium or Chrome, and X11 packages. Debian and Ubuntu can receive system packages through apt.
Start the display with start-display.sh, then copy the glovebox block from mcp-config.json into the MCP client configuration. macOS and Windows are not supported because the sandbox depends on Xephyr. Browser uploads should use upload_file, since the nested file picker is not suitable for automation.
What it helps with
- keep agent actions inside a nested X11 display
- drive browser controls from screenshots
- enter Unicode text into desktop apps
- upload files through Chrome DevTools Protocol
- run more than one isolated app instance
Example prompts for the agent
- Start
start-display.sh, open the target site, and capture its current state with thescreenshottool. - Use
parse_screento find the submit control, then activate it withclick_element. - Send
report.pdfthroughupload_fileinstead of opening the browser’s file picker.
How to install glovebox-mcp
Install
The MCP registry has no launch command for it: see the repository README
Step by step, with commands for each platform: How to add an MCP server to Claude Code, Codex, Cursor, VS Code, Gemini CLI and more
Questions about glovebox-mcp
How do I install glovebox-mcp in Claude Code?
Clone the repository and choose the none mode with git clone https://github.com/segentic-lab/glovebox-mcp && cd glovebox-mcp && ./install.sh none. Then copy the glovebox block from the generated mcp-config.json into Claude Code’s MCP configuration and restart the client.
What does glovebox-mcp do?
It gives an AI agent MCP tools for controlling a browser or desktop application inside a Xephyr window. The actions stay on display :1 rather than the host desktop.
What does glovebox-mcp need to run?
It runs on Linux with Python 3.10+, uv, Chromium or Chrome, and the listed X11 packages. The local vision mode also needs an NVIDIA GPU with at least 6 GB of VRAM.
