Kube Check

Agents write Kubernetes manifests from memory, and memory is several versions old: extensions/v1beta1 Ingresses (gone since 1.22), batch/v1beta1 CronJobs (gone since 1.25), imagePullPolicy: always, memory: 512mb, a selector that does not match the pod labels. The cluster refuses them at deploy time, or worse, accepts them and runs something else. Kube Check checks manifests against the API of the Kubernetes version you run or are upgrading to, without a cluster:

• APIs: what that version serves, what was removed (with the replacement, and the other fields that must change: an Ingress moving to networking.k8s.io/v1 also needs pathType and backend.service), and what is deprecated there, for 1.19 to the newest release.

• Fields: unknown or misspelt fields with the one meant, wrong types, enumerated values the OpenAPI document leaves out (Always, not always), quantities (512Mi, not 512mb), and what the API server refuses beyond the schema: names, labels, selectors that match no pods, requests above limits, probes with two handlers, mounts of volumes that do not exist.

• Custom resources: checked against their CRD when it is in the files, else the CRDs catalog (cert-manager, Argo, Istio, Prometheus and hundreds more).

• Pod Security Standards: baseline by default, restricted on request, check by check as Pod Security admission applies them; plus risks such as unpinned images, missing requests and secrets written into env.

Every finding has its file, line, rule and fix. Look up any field's meaning, or which apiVersion a kind needs, in the same version. No cluster or key needed.

Built and maintained by Arhan Canli.

Install

Needs Node.js 20 or newer. No account or key.

Claude Code

claude mcp add kube-check -- npx -y kube-check-mcp

Claude Desktop: download kube-check-mcp-<version>.mcpb from the latest release and open it. The bundle is signed; verify it with gh attestation verify <file> --repo arhancanli/kube-check-mcp.

Any other client (Windsurf, Zed, Cline, Continue and others), in its MCP config file:

{
  "mcpServers": {
    "kube-check": {
      "command": "npx",
      "args": [
        "-y",
        "kube-check-mcp"
      ]
    }
  }
}

Docker

docker build -t kube-check-mcp https://github.com/arhancanli/kube-check-mcp.git && docker run -i --rm kube-check-mcp

Hosted (Streamable HTTP): node src/server.mjs --http serves stateless MCP at POST /mcp (port from PORT, default 3000).

Related MCP server: k8s-traps

Example

An agent calls check_manifests with:

{
  "files": [
    {
      "path": "k8s/app.yaml",
      "content": "apiVersion: apps/v1\nkind: Deployment\nmetadata:\n  name: Web_App\n  labels:\n    app: web\nspec:\n  replicas: 2\n  selector:\n    matchLabels:\n      app: web\n  template:\n    metadata:\n      labels:\n        app: website\n    spec:\n      containers:\n        - name: web\n          image: nginx\n          imagePullPolicy: always\n          port:\n            - containerPort: 80\n          resources:\n            requests:\n              memory: 512mb\n              cpu: \"2\"\n            limits:\n              cpu: \"1\"\n          env:\n            - name: DB_PASSWORD\n              value: hunter2\n            - name: DEBUG\n              value: yes\n          securityContext:\n            privileged: true\n---\napiVersion: extensions/v1beta1\nkind: Ingress\nmetadata:\n  name: web\nspec:\n  rules:\n    - host: example.com\n      http:\n        paths:\n          - path: /\n            backend:\n              serviceName: web\n              servicePort: 80\n---\napiVersion: v1\nkind: Service\nmetadata:\n  name: web\nspec:\n  selector:\n    app: web\n  ports:\n    - port: 80\n      targetPort: http\n---\napiVersion: batch/v1beta1\nkind: CronJob\nmetadata:\n  name: cleanup\nspec:\n  schedule: \"0 3 * *\"\n  jobTemplate:\n    spec:\n      template:\n        spec:\n          containers:\n            - name: c\n              image: busybox:1.36\n---\napiVersion: policy/v1beta1\nkind: PodSecurityPolicy\nmetadata:\n  name: restricted\nspec: {}\n"
    }
  ]
}

and gets back (recorded from the live server on 2026-09-27):

{
  "kubernetes_version": "1.37",
  "objects": 5,
  "counts": {
    "error": 12,
    "warning": 5
  },
  "findings": [
    {
      "file": "k8s/app.yaml",
      "line": 4,
      "object": "Deployment/Web_App",
      "severity": "error",
      "rule": "invalid-name",
      "path": "metadata.name",
      "message": "The Deployment name \"Web_App\" is invalid: lowercase letters, digits, '-' and '.', starting and ending with a letter or digit."
    },
    {
      "file": "k8s/app.yaml",
      "line": 9,
      "object": "Deployment/Web_App",
      "severity": "error",
      "rule": "selector-mismatch",
      "path": "spec.selector",
      "message": "The selector does not match the pod template's labels (app=website): the API server refuses it.",
      "fix": "make spec.template.metadata.labels contain every selector label"
    },
    {
      "file": "k8s/app.yaml",
      "line": 20,
      "object": "Deployment/Web_App",
      "severity": "error",
      "rule": "schema",
      "path": "spec.template.spec.containers[0].imagePullPolicy",
      "message": "\"always\" is not an allowed value",
      "did_you_mean": "Always",
      "allowed": [
        "Always",
        "IfNotPresent",
        "Never"
      ]
    },
    {
      "file": "k8s/app.yaml",
      "line": 21,
      "object": "Deployment/Web_App",
      "severity": "error",
      "rule": "schema",
      "path": "spec.template.spec.containers[0].port",
      "message": "unknown property \"port\"",
      "did_you_mean": "ports"
    },
    {
      "file": "k8s/app.yaml",
      "line": 25,
      "object": "Deployment/Web_App",
      "severity": "error",
      "rule": "schema",
      "path": "spec.template.spec.containers[0].resources.requests.memory",
      "message": "\"512mb\" is not a quantity: write 512Mi (powers of 1024) or 512M (powers of 1000); units are Ki Mi Gi Ti Pi Ei or k M G T P E, and m means thousandths",
... (127 more lines)

Tools

How it behaves

• Read-only: no tool changes anything outside this process. Manifests never leave it; only schemas and version data are downloaded.

• Network: HTTPS only, to raw.githubusercontent.com and endoflife.date, with a deadline, a size cap and bounded retries. Sources: each version's API definitions from kubernetes-json-schema (generated from Kubernetes' own OpenAPI document; one 1.5 MB file per version, fetched once and kept), removal and deprecation history from Pluto (Apache-2.0), custom resource schemas from the CRDs catalog (MIT), and the release list from endoflife.date. Nothing is logged except unexpected failures (to stderr, without your inputs).

• Read as kubectl reads: YAML 1.1 (yes is a boolean, so value: yes in env is refused, as the API server refuses it), every document of a file, List objects split, comment-only documents (Helm's disabled templates) skipped, null fields treated as unset.

• Whether an apiVersion is served comes from that version's own definitions; Pluto's table only adds when it was deprecated and removed, and a deprecation is reported only when the version also serves a newer API to move to. A field the CRDs catalog's schema does not know is a warning, not an error, unless it is a near miss of a known field: the installed CRD may be newer.

• Patterns written for Go's regular expressions ((?i), \z, [[:alpha:]]) are translated; one that still cannot run in JavaScript is not checked rather than failing the whole schema.

• Results are compact JSON with a matching output schema: errors first, then warnings, then notes; at most 200 findings are listed and the rest counted.

Benchmark

Not yet measured.

Performance

Measured 2026-09-27 from Dubai, home connection against the live upstream, Node 24.19.0 (bench/perf.json, scripts/perf.mjs in the factory).

First call: a fresh server process, including the TLS connection and the upstream's own time. Repeat: the same call again, answered from the in-process cache, so it shows this server's own overhead.

Tool definitions the model reads on every turn (name, description, input schema): 2,212 characters. The full tool list, with the output schemas and annotations clients use to validate results, is 3,614 characters.

More MCP servers by Arhan Canli

• Actions Check: Checks GitHub Actions workflows: outdated actions, old Node runtimes, retired runners, injection.

• Config Check: Validates config files against their official schemas: tsconfig, compose, workflows, 1,400+ more.

• Cron Check: Explains cron expressions, lists next run times in any time zone, converts between cron dialects.

• Dockerfile Check: Checks Dockerfiles: build-breaking mistakes, base image tags that exist, digests, platforms, EOL.

• Domain Health: Email and domain checks: SPF lookup limits, DKIM keys, DMARC, DNS records, registration expiry.

• End of Life: Is this version still supported? EOL dates, latest patch and upgrade target for 470+ products.

• Internet Standards: RFC sections, status, obsoleted-by chains, errata and IANA registries for coding agents.

• License Check: Open source license answers: SPDX ids, copyleft, and whether a dependency's license fits yours.

• The whole collection, 12 more

License

MIT, Copyright (c) 2026 Arhan Canli.