tastyeffectco/sandboxes

Sandboxed is an open-source backend for building products like Lovable, Bolt, or Replit, where a user types a description and a working web app appears at its own URL seconds later. If you want to build that kind of product yourself, this is the infrastructure layer that makes it work. When you send it an API request, it creates a private isolated container for that user, runs an AI coding agent inside it with a prompt you provide, and exposes the running app at a shareable preview URL. Each container is isolated so one user's code cannot see or interfere with another user's. The AI coding agents (OpenCode and Claude Code) are pre-installed in every container, so you just send a text prompt and the agent writes and runs the code. A key design choice is cost efficiency. Containers go to sleep when no one is using them, freeing memory, and wake up automatically when someone opens the preview link. This means one ordinary server can host many users at once instead of needing a dedicated virtual machine per user. The tool handles the wakeup timing, shows a loading page while the container restarts, and resumes serving traffic once it is ready. The technical stack is deliberately simple: one Go binary as the control plane, Docker for running containers, Traefik for routing preview URLs and handling HTTPS, and SQLite as the database. There is no Kubernetes requirement and no separate database server. On a reboot the system reads SQLite and reconciles the running containers back to the expected state. You install it by cloning the repository and running a single install script on a Linux machine that has Docker. The API is then available at a local port and accepts requests to create sandboxes, run agent tasks, stream progress, and manage container lifecycles. The project is aimed at teams building multi-user AI coding tools or preview environments, not at developers who just want to run one or two containers for themselves. It is released under the MIT license and is described as beta.