Playbooks

The base of the platform: unmodified upstream Kubernetes on your machines, with identity, storage, registry, database, certificates and GPUs installed and connected before you open the first page, and Thinkube Control to drive it all from the agent.

Ask your agent: "what’s running?" The agent lists every service of the base and everything you added: Thinkube Identity, Thinkube Storage and JuiceFS, Thinkube Registry, Thinkube Git, Argo Workflows and ArgoCD, PostgreSQL, the GPU Operator, Thinkube IDE, Thinkube Notebooks, Thinkube Experiments, Thinkube Packages, your optional components and your apps.

How it works

  • One sign-on. Thinkube Identity, powered by Keycloak, is the login of every service, every dashboard and every app you deploy. You sign in once.

  • One filesystem, one registry. Thinkube Storage is S3-compatible object storage, and JuiceFS puts a shared filesystem on it: the IDE workspace, the notebooks, the datasets and the model weights are one place. Thinkube Registry, powered by Harbor, holds your app images, the model servers and the notebook images.

  • Every service at <name>.<your domain>, from any device on your tailnet (your private Tailscale network). Every machine and the cluster’s gateway join your tailnet, and your domain resolves on that network only. The cluster’s DNS answers every name under your domain with the gateway’s address, so a new app needs no DNS record. The gateway holds one wildcard certificate from Let’s Encrypt, a certificate that renews itself, and routes each name to its service. The services are reachable on your tailnet only.

  • A machine that moves keeps working. When a machine gets a new network address, for example from a new router, the cluster keeps working within a minute.

  • GPUs as slots, memory as a budget. The GPU Operator shares each GPU as four slots, so a served model, a fine-tune and an embedding server run on one card. A vLLM, TensorRT-LLM or text-embeddings model takes one slot; Ollama holds several models in one. The platform counts memory per node: on a card, the card’s memory; on a DGX Spark, where GPU memory is host memory, a budget of 96 GB (112 GB on a node dedicated to AI) that keeps the machine responsive. A load starts only when it fits.

  • amd64 and arm64 in one cluster. Apps run on both amd64 and arm64 machines in one cluster. Images and notebook environments are built for each architecture, and each pod runs on a machine that can run it.

  • Every operation is a sentence. Thinkube Control exposes the same operations to its pages and to the agent. The agent reads state freely and changes it when you ask; What Thinkube is says where the line is drawn.

Playbooks

Thinkube Kubernetes 30 min

Check and restart services from your agent

beginner

Ask what is running and whether it is healthy, restart a service, and turn one off and on again, from a chat with your agent

2026-10-04 Thinkube Kubernetes 30 min

Install and remove optional components

beginner

Add a database, a vector store or an inference backend to the platform when you need it, and take it out again

2026-10-04

Reference