
kube-noob
io.github.kube-noob/kube-noob
Kubernetes, OpenShift, ROSA HCP, Helm and RHACS audit, debugging and cited docs. No AI inside.
Documentation
kube-noob
No AI inside. Works with yours. kube-noob is a Kubernetes, OpenShift, ROSA HCP, Helm and RHACS/StackRox expert in a box: audit rules, read-only-first playbooks, cited docs and graded exercises. Claude, OpenShift Lightspeed or any MCP client calls it as tools; the CLI runs it with no AI at all. No network calls unless enabled. Browse the rules, the demos and the recordings at kube-noob.com.
Two questions it answers that other tools do not, offline, from files you already have:
kube-noob policy reach --from web/frontend --to shop/payments --port 9090 -f examples/netpol-admin-tiers.yaml
kube-noob policy explain -f examples/denial-multi.txt
Which policy governs this traffic? AdminNetworkPolicy → NetworkPolicy → BaselineAdminNetworkPolicy, in
cluster order, naming the tier, policy and rule that decided. A checker reading NetworkPolicy alone gives a
wrong answer wherever a platform team set cluster-wide guardrails, not a narrower one.
Which system denied this pod? One pass across Pod Security, SCC, Kyverno, Gatekeeper, VAP and RHACS, with the
smallest fix. reach exits 1 when denied; both say what they could not determine.
Also: 93 rules with strict schemas and 9 red-team scenarios; 67 offline playbooks, read-only steps first;
must-gathers triaged into a case draft that says what could not be read; OpenShift bugs with fix versions; hardened
manifests and admission policies; 5 tracks, 16 lessons. kube-noob gate puts any of it in CI with SARIF and one
exception list, and a design request runs the 7-step protocol (Clarify, Validate, Research, Audit, Calibrate,
Generate, Red Team). Answers cite local knowledge as [kn:<doc>#<n>]. All of it: use-cases.md.
Who does the thinking
| You use | Who answers | What leaves your machine |
|---|---|---|
| The Claude plugin, or any MCP client | That client's model, calling kube-noob's tools | What you send that client; kube-noob sends nothing |
| The CLI | kube-noob's rules, no model | Nothing |
kube-noob ask (optional) | A model you configure | Your question, to that model only |
Quick start
brew tap kube-noob/kube-noob https://github.com/aashishmahajan/kube-noob
brew trust --formula kube-noob/kube-noob/kube-noob
brew install kube-noob
kube-noob demo --list
kube-noob audit -f your-manifests/ --platform openshift --version 4.22
kube-noob insight 'paste anything'
kube-noob doctor
No Homebrew: uv tool install kube-noob (Python 3.11.4+); install help.
Use with Claude or any MCP client
Claude plugin, in Claude Code (needs uv):
/plugin marketplace add aashishmahajan/kube-noob
/plugin install kube-noob@kube-noob
Desktop app, claude.ai chat and the Agent Skill: install.md. Any MCP client:
{"mcpServers": {"kube-noob": {"command": "uvx", "args": ["--from", "kube-noob==0.9.2", "kube-noob", "mcp"]}}}
Run on OpenShift
As a token-protected service, loopback-only sidecar or Lightspeed tool:
docs/openshift.md. The chart ships
NetworkPolicies, written for restricted-v2.
Safe by default
Off until you turn it on: web lookups, live cluster access (read-only, never Secrets), and any model at all. An
admin offline: true lock overrides every one of them and nothing later in the config chain can lift it. Config
never comes from the current directory. kube-noob doctor shows what the process can reach and exits 1 if the
configuration contradicts the lock; every key is in kube-noob.example.yaml, and the
generated safety contract lists what each capability tier can do.
Docs
Everything is at kube-noob.com, and in docs/: use cases, architecture, accuracy, guidance, learning, privacy, security, contributing, limits, roadmap, maintenance, governance.
Credits
Notices: THIRD_PARTY_NOTICES.md. Kubernetes, Helm, OpenShift, ROSA, RHACS, StackRox, OpenShift Lightspeed, Claude, Anthropic, GitHub and AWS are trademarks of their respective owners; kube-noob is independent and not endorsed by them. License: Apache-2.0.
kube-noobpip install kube-noob