sandboxed is an open-source platform that creates self-hosted development environments with live preview URLs through a single command. The tool generates isolated Linux containers for each user, installs AI coding agents like OpenCode and Claude Code, and automatically provisions shareable preview links. Built for AI app-builder products, it eliminates the need for Kubernetes while providing multi-tenant isolation and cost-effective resource management.

The system runs on Docker with Traefik handling URL routing and SQLite managing state, all controlled by a lightweight Go binary. When sandboxes idle, they sleep to free memory and wake instantly when accessed, allowing dozens of environments on modest hardware—a $2,000 server can host many users. The architecture uses hardened containers with security restrictions and persistent workspace storage.

Developers interact with sandboxed through a simple HTTP API that creates containers, executes commands, and streams AI agent progress via Server-Sent Events. The first request to a stopped sandbox automatically wakes it, and preview URLs resolve locally without requiring DNS setup. Users can inject their own API keys for custom model access while the base image includes free-tier agent tools.

This solves the infrastructure complexity behind AI-powered coding platforms where users describe what they want built and receive working applications instantly. By distilling months of platform engineering into one command—`./install.sh`—sandboxed makes self-hosted AI development environments accessible to any team with Docker installed.