Hi, I'm Abdelouahed, building sandboxd in Zurich

Hi, I'm Abdelouahed, a founder in Zurich building sandboxd, an open-source, self-hosted AI app builder. You describe an app, and a coding agent builds it in an isolated container that appears at its own live preview URL. You can also run any of 80+ open-source apps the same way. It installs with one command and needs only Docker, with no Kubernetes or database server. The code is MIT-licensed on GitHub, where it has around 970 stars.

The part I'm most particular about is that no credentials live in the sandbox. A proxy injects model keys on the wire, and every agent task is checkpointed first, so any change can be undone in one click. My stack includes Claude, Claude Code, Codex, ChatGPT, DeepSeek, Langfuse for tracing, and Apollo.io for outreach.

What I'd love from this community is critique from anyone who has run agent-generated code on their own infrastructure, especially on isolation and idle/wake behavior. If you've built preview routing or hardened sandboxes, I'd value your read on the architecture guide at sandboxd.io.