litek8s

A whole Kubernetes in one binary.
A local cluster in about a second, your spare machines as one cluster,
and a control plane that idles at 23 MiB.

Website · Docs · Quickstart · Releases

brew install jairoFernandez/tap/litek8s
litek8s dev up        # a cluster, Ready in ~1 s
kubectl get nodes     # your kubectl, Helm and Argo CD work unchanged

Why litek8s

• A cluster before you finish blinking. litek8s dev up has a node
  Ready in 0.17 s on Linux and 1.28 s on macOS; kind takes 31.6 s and 34.3 s.
• Small enough to forget it's running. The whole control plane idles at
  23 MiB where k3s needs 567 MiB, 33.5 MiB with five nginx pods (k3s: 579);
  with Argo CD, 90 MiB against 900 MiB. On macOS a dev cluster runs in a
  small Linux VM and Activity Monitor shows the whole VM: about 240 MiB
  fresh, more once it has pulled and run images.
• Real Kubernetes, not a look-alike. It serves the Kubernetes API, so
  kubectl, Helm and Argo CD just work: 119/119 on our compatibility corpus
  (47 Helm charts included) and 419/419 on the official v1.33 conformance suite.
• Batteries included. A Docker-compatible API, a terminal UI, metrics,
  log search and alerts (+1.8 MiB, instead of 1.3 GiB for
  kube-prometheus-stack and Loki) and an MCP server for AI agents, all in
  the same binary.
• No Docker Desktop, no VM to babysit. On macOS one small VM it manages
  itself, no sudo; on Linux a rootless sandbox.
• From laptop to rack. The same binary turns a Raspberry Pi, an old
  laptop and a cheap cloud VM into one cluster, across NATs, with a highly
  available control plane when you want one.

_{litek8s ui: your clusters, pods, containers, images and logs in one keyboard-driven terminal UI.}

What you can do

• Keep the clusters you have. litek8s only adds contexts to your
  kubeconfig, next to EKS, GKE or Civo; same manifests, same charts, and
  going back is deleting it. → Next to your clusters
• Develop locally. As many clusters as you like, each with as many
  nodes, NodePorts on localhost and your code mounted at the same path.
  → Dev clusters
• Drop Docker Desktop. docker, docker compose, Dev Containers and
  Testcontainers run on the same cluster through its Docker Engine API.
  → Docker API
• Kubernetes in CI: the fastest setup step. One GitHub Actions step
  gives the job a cluster (and a Docker API) Ready in 2.5 s instead of
  kind's 45 s; Helm and multi-node jobs finish 3–5× sooner with 26–37× less
  memory, and 1.8× sooner than kind in Docker in Docker on GitLab-style
  container jobs. → CI
• Turn spare machines into a cluster. Install a server and its nodes
  over SSH, or declare the whole cluster in one file and run
  litek8s cluster apply -f cluster.yaml.
  → Clusters over SSH
• Span homes and clouds. Pods talk across nodes over a WireGuard mesh
  litek8s sets up itself, relayed when two machines sit behind NATs: no VPN,
  no port forwarding. → Networking
• Run one LLM across all your machines. A model too big for any one of
  them runs on several with llama.cpp's RPC backend, as a DaemonSet behind
  one OpenAI-compatible endpoint. → One LLM across machines
• Give pods your GPUs. A node's CDI specs (nvidia-ctk cdi generate)
  turn into nvidia.com/gpu (or amd.com/gpu...) resources pods ask for in
  their limits: no device plugin, no operator. → GPUs
• Stay up when a server dies. Three API servers share the store through
  Raft, no etcd; any one can fail without losing a write.
  → HA control plane
• Let AI agents operate it. litek8s mcp gives Claude Code and other
  MCP clients your clusters, pod diagnosis, logs and metrics, with
  guardrails. → Agents
• See what's happening. Prometheus-compatible metrics (Grafana works),
  log search across every pod and alerts to a webhook, ntfy or Telegram.
  → Observability

Install

macOS (Apple silicon) and Linux (x86_64, aarch64):

brew install jairoFernandez/tap/litek8s

or

curl -fsSL https://raw.githubusercontent.com/jairoFernandez/litek8s/main/install.sh | sh

On Linux, litek8s dev setup checks what rootless clusters need. A server
for your other machines is one command too: install.sh | sh -s -- server.
Other options (cargo, releases, servers and nodes):
Install.

How it compares

Measured with the scripts in this repository; setups, raw data and caveats
in Numbers.

Limitations

litek8s is an early preview (v0.x), built and tested by one person.

• Authentication and RBAC are simpler than upstream's (one OIDC issuer at most, no
  Node authorizer, no certificate revocation). Servers enforce RBAC
  (--authorization-mode RBAC-audit only logs); dev clusters listen on
  127.0.0.1 only and every client there is admin.
• APIs, flags and on-disk formats may change between 0.x releases.
• Some kubelet behaviors are missing.

The full list: Limitations.
Feedback and issues are very welcome.

Learn more

• litek8s.dev: the tour
• Documentation: next to your clusters, how it works,
  platforms, multi-node and HA,
  CRDs
• Contributing · Security

History

litek8s began as a small Docker-backed orchestrator with its own daemon and
manifest format, plus an experiment in distributed LLM inference. 0.2.0
removed both; every command they had has a Kubernetes-native replacement,
listed in CHANGELOG.md.

License

Apache-2.0, see LICENSE and NOTICE. Third-party
components: THIRD-PARTY.md.