blog.moebius01.cc

MCP Infrastructure Agent

An MCP server exposing infrastructure operations — Proxmox VM lifecycle, Ansible playbook runs, Wazuh queries — as tools an agent can call. Authorised, audited, opt-in.

Most infra work is small, repetitive, and well-defined: list VMs, restart a service, pull a Wazuh alert, run a hardening playbook. The Model Context Protocol gives us a clean way to expose those operations to an LLM with explicit authorisation and an audit trail — instead of giving the model shell access.

Design constraints:

  • Read-only tools by default, write tools require human confirmation.
  • Every call logged with the model identity, prompt context, and outcome.
  • Scoped tokens per environment (lab vs prod), never one keyring.

Status: Architecture sketched, building the read-only Proxmox + Wazuh tools first. Write paths come behind a feature flag and a review.