zerobox

Why we included this project

Zerobox runs commands from AI agents inside a sandbox that blocks writes, network access, and environment variables by default, so nothing happens unless you explicitly allow it. It can also inject real API credentials for approved hosts only, which means the process never sees the keys themselves. That is exactly what teams wiring LLM tools, MCP servers, or coding agents into their workflows need: a hard boundary between what the model can do and what it can reach. The whole thing ships as a single binary with Rust, TypeScript, and Python SDKs, so you can add it to an existing pipeline without spinning up Docker or a VM. If you need to contain agent-generated commands, this is a lightweight option worth a look.