SkillPartly tested (needs an account)

Kubernetes ops gap audit

kubegap

A read-only Kubernetes audit skill for Claude Code and Codex. It runs 411 checks on your cluster and writes a plain report with evidence and fixes.

Works with
Claude Code, Codex CLI
Good for
Security testing

What it does

kubegap is a skill that turns your coding assistant into a read-only Kubernetes operations auditor. You point it at a kube context you already use. Scripts collect the cluster facts and judge 409 of its 411 checks in 12 areas. The agent then picks the problems that matter now and writes a plain report: what is wrong, why, how to fix it and how to roll the fix back, with the evidence.

The daily audit is the short routine and lists 3 problems. The full audit reads more (inside the nodes, control-plane logs, extra tools, and on EKS cost and ECR) and lists 5. It is not a monitoring tool and it does not fix anything for you. It finds the gaps in your cluster and in the monitoring you already run. EKS and k3s get the deepest coverage; GKE, AKS and others get the in-cluster checks.

Example prompts

  • "/kubegap"
  • "/kubegap full"
  • "/kubegap full focus security"
  • "/kubegap explain <finding id> --snapshot latest"
  • "/kubegap --context <name>"
  • "/kubegap setup"

How to install

npx skills add KyungHwanKim-devs/kubegap -g

How to use it

In Claude Code type /kubegap for a daily audit, or /kubegap full for a full audit. In Codex type $kubegap. It asks the report language (English or Korean) and confirms the kube context, may ask first-run setup questions, then writes report.md and report-details.md.

Check it works

After install, open Claude Code or Codex on a machine that has kubectl and python3 3.11 or newer. Type /kubegap (Claude Code) or $kubegap (Codex). It should first ask which language you want for the report, then ask whether to audit the current kube context. Answer No if that is the wrong cluster. At the end it gives you the paths of report.md and report-details.md.

Common problems

Other assistants

What to watch out for

  • Needs kubectl with read access to the cluster, python3 3.11 or newer and jq; on EKS also the AWS CLI. The npx install line needs Node.js 22.20 or newer
  • Read-only on the cluster by design: SKILL.md forbids kubectl apply/delete/scale, aws update/put/delete and helm upgrade. The one Secret read is helm list in a full audit (Helm release records, metadata only)
  • Cluster data goes to your AI provider: the agent reads the snapshot files (summary, findings, raw files it opens) and that goes to Anthropic or OpenAI. docs/DATA.md says there is no mode that keeps the data local
  • Report writing starts up to four more headless agents with claude -p or codex exec, on your login and plan usage
  • First-run choices can install tools on your machine (krr, polaris, pluto, kubent, helm, aws, jq): Homebrew when present, else pinned, sha256-checked downloads into the skill's .tools/ folder. Only after you pick Install
  • On EKS, the Upload choice creates the SSM document kubegap-node-read in your AWS account (needs ssm:CreateDocument); node reads then run as root through SSM. On k3s and self-installed clusters it uses SSH with your own keys after you choose it. If node access is already set up, it reads the nodes without asking
  • Stores snapshots under ~/.kubegap/ (about 40 MB per full audit); prune with scripts/kubegap prune. Config values are written with scripts/kubegap config set only after you confirm
  • While it writes the report, it may open the public incident links and official docs behind a problem with the agent's web tool
  • Each snapshot format change makes kubegap refuse older snapshots (0.1.5 and 0.2.0 both do this). After an update, run a new audit before you use --compare or explain on older audits
  • No disable-model-invocation, and the description invites use for "audit my cluster", so the agent can start it without the slash command

Licence and author

Licence
MIT
Author
KyungHwanKim-devs
Source code
github.com/KyungHwanKim-devs/kubegap

How we tested it

We installed it on an empty test computer on 7 Nov 2026. Partly tested: installed in a clean test system, but its main features need an account or key we did not use.

Test level

Not run against a real cluster. The Claude Code skills docs list only ~/.claude/skills and .claude/skills as skill folders, so a folder only in ~/.agents/skills is not loaded by Claude Code.

Technical details of the test
Date
7 Nov 2026
Method
Reviewed against the repository at 787d301 (v0.1.5): README, SKILL.md, docs/SETUP.md, docs/DATA.md, docs/HEADLESS.md and lib/tools.py, lib/writer.py. Install tested in a clean container (oexp-sandbox image, no account) with -y added to the README line to skip the prompts: npx skills add KyungHwanKim-devs/kubegap -g -y put the folder only in ~/.agents/skills; with -a claude-code also added it landed in ~/.claude/skills/kubegap and scripts/kubegap is executable.

Checked again on 8 Nov 2026

Alternatives

No other catalog entry audits a running cluster. The closest:

  • OpenQodex — scans the Kubernetes files in your unpushed changes, not a live cluster
  • CrowdSec Skill — helps install and fix CrowdSec (also on Kubernetes) to block attacks; it does not audit cluster operations

FAQ

  • Q: Does it change my cluster? A: The README and SKILL.md say it only reads the cluster and installs nothing in it. Writes happen on your machine (~/.kubegap/) and, if you choose Upload on EKS, one SSM document in your AWS account.
  • Q: Does my cluster data leave my machine? A: Yes, to your AI provider. What the agent reads goes to Anthropic (Claude Code) or OpenAI (Codex). Secret values are not collected, and reports are masked, but the snapshot itself is not.
  • Q: Do I need Prometheus? A: No. If a monitoring tool is missing, its checks skip and the report says what it could not see instead of counting a pass.
  • Q: Is it only for EKS? A: No. EKS and k3s have the deepest coverage. GKE, AKS and kubeadm get the in-cluster checks; nodes are read only where SSM or SSH works.

Sources