An AI-operated infrastructure control node
How Claude built and now runs a production Linux server end to end — no manual SSH intervention.
The premise
Michael, a Systems Engineer who runs infrastructure for a managed-services company, wanted to test a real question: given direct, authenticated, audited access to a server, can an AI actually operate it — not just answer questions about it, but do the work?
What was built
- An MCP (Model Context Protocol) server exposing controlled tools — run commands, manage services, read/write files, check system state — over an authenticated, TLS-terminated HTTPS endpoint
- Every single call audit-logged: tool name, caller IP, timestamp, and parameters, with 90-day retention
- A three-layer firewall (hypervisor, host, and application-level IP allowlisting) protecting the control endpoint
- SSH hardened to key-only auth with no password login
- A dedicated 200GB storage volume provisioned and mounted
- A full multi-tenant web hosting stack — nginx, per-site Linux users, per-site TLS certificates, optional password protection — built as a repeatable pattern, deliberately isolated onto a separate IP from the control plane so the two never interfere
- This very site (and the one you're reading this on) were both provisioned by that pattern
What actually happened along the way
Not everything worked on the first try — and that's the point of showing this instead of just claiming it:
- A port conflict was diagnosed with tcpdump and packet capture across multiple network layers before the fix (moving a service off a non-standard port) was identified
- A new storage volume came up with an unrecognized SELinux security context, silently blocking the web server until diagnosed via audit logs and fixed with a persistent policy rule
- A misconfigured default web server binding was caught and corrected before it could interfere with an unrelated certificate-renewal process
- Public reachability for a new site was debugged layer by layer — DNS, edge proxy, hypervisor firewall, host firewall — using the specific error codes returned at each stage to isolate exactly where the block was
Where it stands
The control node is live and in daily use. Full technical documentation, kept current as the system evolves, is public at sentinel.aod.net/docs (password-protected).