MCP servers have a way of multiplying. One for GitLab, one for Grafana, one for Vault, one for the cluster. Before long each AI tool has its own list, half of them run as local processes, and the config files are full of long-lived tokens.

This series follows how we turn that into a platform. The MCP servers move into Kubernetes, behind one internal gateway that every AI client in the house can use: Claude Code, Codex, Hermes Agent, our self-hosted open-webui and n8n, and Home Assistant. People sign in through Keycloak. Workloads prove who they are with SPIFFE. No client keeps a static secret, and no MCP server is reachable except through the gateway.

The parts are written as the work happens, so each one describes something we actually built and ran, including what didn’t work.

What the series covers #

  • The design: what each client can really do, why there’s no public endpoint, and what SPIFFE does and doesn’t solve.
  • SPIRE everywhere we need it, including a self-built agent for macOS.
  • The gateway: two token issuers, policy per caller and per tool, mTLS to every MCP server.
  • Backend credentials: Vault login with a SPIFFE identity, short-lived Kubernetes tokens, and what to do with APIs that only accept static keys.
  • Migration: from MCP servers on a Docker host to namespaces on Talos, dev first, then production.

Before you start #

Expect a Kubernetes cluster, a Vault you control and an OpenID Connect provider such as Keycloak. The parts point out where our setup is specific to us.