Screenbox: Virtual Desktops for AI Agents
Dmytro Klymentiev
Published
Self-hosted platform that gives each AI agent its own isolated Linux desktop with a real Chromium browser. Agents see the screen, click, type, navigate tabs. You watch them work in real time via built-in dashboard and take control when needed.
Docker-based, MCP-native. Works with Claude Code, Claude Desktop, Cursor, or any MCP-compatible client. Each desktop is a container with full isolation, no bind mounts or host access.
Features: live screenshot streaming, Chrome DevTools Protocol semantics for faster element interaction, snapshot and restore of desktop state, knowledge compilation from past sessions, human-in-the-loop control.
Around 2GB RAM per desktop, no GPU required. Runs on any Linux server or Windows with WSL2.
Site: https://screenbox.dev
GitHub: https://github.com/dklymentiev/screenbox
Demo: