Kubexa

work in progressThis site is being written now: pages appear and change as the cluster does. The AI layer is the most incomplete part.

talos linux · flux cd · git as the only interface

One cluster.
Everything in git.

Four machines, one repository, nothing applied by hand: identity, secrets, CI, storage, networking and the applications on top. Every page quotes the manifest that runs it.

Read the docs/llms.txt

The stack, bottom up

Eight layers, in the order they had to exist.

01Talos LinuxNodes boot an immutable image and take their configuration from a machine config. No SSH, no package manager, no drift.taloslinuxkubernetes1 page02HardwareTwo physical machines with consumer CPUs, one OCI VM, one libvirt VM. The specs decide what can be scheduled where.nvidialinuxkubernetes1 page03GitOpsOne repository, reconciled by Flux every few minutes. Renovate opens the pull requests, image automation rewrites the tag, and nothing else changes the cluster.fluxkubernetesrenovatehelmharbor1 page04NetworkCilium is the CNI and the gateway. Gateway API resources define every hostname, external-dns writes the records, and two mesh networks cover the tailnet and the two sites.ciliumcloudflaretailscalewireguard1 page05StorageCeph replicates each object across three nodes and hands out volumes; CloudNativePG runs PostgreSQL on those volumes with backups to object storage.cephpostgresqlcloudnativepg1 page06MonitoringMetrics, logs and traces with a short list of alerts. A signal earns its place only if it changes what I do at 3am.victoriametricsgrafanaprometheus1 page07wipAI stackWork in progress. Sandboxed agent devboxes and MCP servers run today; model serving and GPU sharing are still moving.kubernetesrustnvidia1 page08Open source projectsCharts, build tools and conventions that came from other people's repositories, plus the projects contributed to from this one.cncfhelm3 pages··Full indexEvery document on one page, grouped by layer.all layers

What a page looks like

The reader itself, not a mockup.

A kubexa documentation page: a contents sidebar beside a page that quotes its HelmRelease

Every page names the manifest it documents and quotes it from the same file the cluster runs.

  • the quote is resolved at build time from the repository
  • a page cannot cite a file that is not there
  • every page also exists as plain Markdown at the same URL plus .md
The hardware page: node CPU, memory, GPU, architecture and siteThe documentation index: eight layers with their pages

What runs right now

Read from the cluster, not from a slide.

nodes
4
3 control-plane, 1 worker, 2 physical
kubernetes
v1.37.1
talos v1.14.4 below it
flux
2.10.0
reconciles its own installation too
pages
10
across 8 layers
gpus
3
RTX 3050 class, one worker

Questions this stack answers

The things that decide whether a platform is operable.

How does a change reach production?

Build, digest, commit, apply, roll: kaniko builds the image, Flux resolves the digest, the image automation commits it, and the Kubernetes objects follow from git. No human touches a tag.

What happens when a node dies?

Ceph re-replicates three ways, CloudNativePG promotes a replica, and Flux schedules the workloads onto the remaining capacity. The next plan shows no drift.

Why quote manifests instead of describing them?

A description drifts the day someone edits the file. These pages resolve their quotes from the repository at build time, so a stale page fails the build instead of lying quietly.

What is still unfinished?

The AI layer. Sandboxes and MCP servers run; model serving and GPU sharing are still moving, and that page says so instead of presenting a settled design.

Start with a layer,
or read it all.

Open the docs

Recent notes

What changed and why.