x-nch/xnchSystems ? reverse-engineered prompt

Build me a private AI orchestration system that feels like a real assistant, but with strict human approval for anything risky.

I want a control plane that handles login, permissions, policy checks, memory, goals, audit logs, and a queue for approving actions. It should talk to a separate decision engine that plans steps, picks workflows, and sends back only approved claims or actions. Include a simple web app where I can chat with it, review memory, see the workflow graph, and approve or reject queued actions. I also want a command line client for everyday use, plus support for voice if that fits naturally.

Keep it fully self hosted, with no cloud inference, and make it work across two machines with local services and systemd based startup, not Kubernetes. Add the plumbing for evals, a docs or MCP style bridge for tools, and enough tests and deployment scripts that I can bring it up end to end and verify it works. If you need to check current docs or best practices, look them up online.