twenty-four: the platform
This is the infrastructure under twenty-four .
The cluster
k3s on hardware that was lying around:
master: Control plane, x86_64arm1,arm2,arm3: Apple Silicon Mac Minis on Fedora Asahi Remix, which do most of the worklab: Old x86_64 desktop, for anything that needs amd64brain: GPU box, switched on when neededgarage-door: A Raspberry Pi that opens the garage
Networking
The router forwards .home DNS to CoreDNS in the cluster, and MetalLB and Traefik expose the services. None of it is on the public internet, so the two services that need to be reachable from outside run on Cloud Run. S3 holds the data, and a Synology NAS provides persistent volumes.
Deploys
Publishing a GitHub Release is the deploy:
- Actions builds multi-arch images and pushes them to GHCR.
- A CronJob in the cluster sees the new release within five minutes and rolls every service to it.
- It waits for the rollout, rolls back on failure, and sends me a notification through signal.
There is no :latest. In September 2026 a deploy defaulted to it, and a stale cached copy on one node ran an old routine against my real calendar for four minutes. The tag checked into each manifest is now :dontexist, so a bare kubectl apply fails instead of guessing.
Credentials
There’s no Anthropic API key in the cluster. Pods authenticate with Workload Identity Federation, using a projected Kubernetes token. Services don’t hold Pushover credentials either; everything goes through signal.
This post was written by Claude, guided & reviewed by Nick.