A new open-source tool called Sandboxd provides self-hosted development sandboxes with preview URLs using Docker and Go, designed specifically as a Kubernetes alternative for teams building AI-powered app builders and agent platforms. Created by tastyeffectco and launched on GitHub in June 2026, the project has already garnered 511 stars and significant community interest on Hacker News with 111 points and 34 comments.
Solving Multi-Tenant Isolation and Preview URL Challenges
Sandboxd addresses a fundamental infrastructure challenge in AI agent development: providing isolated environments where one user's code cannot affect another's, while delivering per-user preview URLs with automatic routing. The project's README emphasizes that "the fundamental challenge isn't crafting prompts—it's the underlying infrastructure." The tool enables three operations in a single HTTP request: container creation with isolated Linux environments, agent execution running OpenCode and Claude Code CLIs inside sandboxes, and live preview delivery through shareable URLs.
Technical Architecture Built for Simplicity
The system architecture deliberately avoids complexity, using a single Go program as the control plane that manages Docker, supplemented by Traefik for routing and SQLite for state management. Sandboxd implements hardened Docker containers with restricted capabilities and read-only root filesystems for security. The tool includes automatic cost optimization by stopping idle sandboxes and transparently waking them on request. Workspace persistence is achieved through bind-mounted directories that survive container lifecycle changes.
One-Command Deployment with Production Features
Sandboxd offers one-command deployment via ./install.sh and requires no separate database servers or message queues. The system generates preview URLs in the format http://s-<id>-3000.preview.localhost with zero DNS configuration required. Production deployments can leverage Let's Encrypt for TLS support. The project promises "dozens of sandboxes sharing one box instead of one VM each," significantly reducing infrastructure costs compared to traditional approaches.
Target Use Cases and Production Readiness
The MIT-licensed tool targets AI app-builders where users describe an app and see it live, agent platforms requiring headless orchestration, per-user preview environments, and SaaS factories needing dense multi-tenant isolation. The creator notes the project is intentionally "simple on purpose"—suitable for startups but requiring additional hardening before scaling to untrusted users in large-scale production environments.
Key Takeaways
- Sandboxd provides Docker-based development sandboxes with preview URLs as a Kubernetes alternative for AI agent platforms, achieving 511 GitHub stars since June 2026
- The system performs container creation, agent execution, and live preview delivery in a single HTTP request with multi-tenant isolation
- Architecture uses one Go program, Traefik routing, and SQLite for state management—eliminating separate databases and message queues
- Automatic cost optimization stops idle sandboxes and wakes them transparently on request, enabling "dozens of sandboxes sharing one box"
- MIT-licensed with one-command deployment suitable for startups, though requiring hardening for large-scale untrusted user environments