Architecture Decision Record

ADR 013: MCP Servers as the AI-Operations Interface

Eleven in-cluster MCP servers give AI tooling structured, scoped access to platform APIs — instead of pasted terminal output or raw credentials.

Status: Accepted  ·  Date: Mar 2026  ·  ← All ADRs


Context

AI assistants became genuinely useful for operating this platform once they could see it — my AI-assisted engineering notes call cluster-specific context the difference between generic answers and correct ones. The question was how to grant that access:

Decision

Eleven MCP servers run in the cluster — Kubernetes, ArgoCD, Gitea, Vault, MinIO, CloudNativePG, Talos, Authentik, Synology, Unifi, and Paperless — each a small Deployment in a dedicated namespace. Design rules:

Reasoning

Tradeoffs

Outcome

Troubleshooting sessions now start with the assistant querying the actual system instead of me transcribing it — the workflow the copilot-experiments page describes as the single biggest context unlock. The MCP layer is here to stay; the Policyclaw prototype stands as the working proof of where the governance goes next: policy-mediated, confirmation-gated AI operations, built into Cortexa as an invariant rather than bolted on as a separate box.