k1c

Experimental kubectl apply-style tool for Cloudflare. Pronounced "kick".

This is a learning / proof-of-concept project. The CLI + operator are both published (@mizchi/k1c on npm, ghcr.io/mizchi/k1c-operator on GHCR), but the public surface is subject to breaking changes without notice.

k1c parses a defined subset of Kubernetes manifests and applies them to a Cloudflare account via the official SDK. The motivation is to reuse kubectl-style declarative manifests for personal-scale Cloudflare projects without paying for a real Kubernetes control plane. Same manifest can be applied either via the CLI (one-shot) or via the operator running inside any k8s cluster (continuous reconciliation, finalizer-driven cascade delete).

apiVersion: cloudflare.k1c.io/v1alpha1
kind: R2Bucket
metadata: { name: media }
spec: { location: weur }
---
apiVersion: apps/v1
kind: Deployment
metadata: { name: api }
spec:
  selector: { matchLabels: { app: api } }
  template:
    metadata: { labels: { app: api } }
    spec:
      containers:
        - name: api
          image: ./dist/worker.js
          volumeMounts:
            - { name: bucket, mountPath: /mnt/media }
      volumes:
        - name: bucket
          csi:
            driver: r2.k1c.io
            volumeAttributes: { bucketRef: media, binding: R2_MEDIA }

$ K1C_ACCOUNT_ID=...  CLOUDFLARE_API_TOKEN=... pnpm k1c apply -f manifest.yaml

-f accepts a path, a directory (every .yaml / .yml recursively; files named _* and .* are skipped), or - for stdin — so the rest of the k8s ecosystem composes naturally:

helm template ./examples/helm-chart | k1c apply -f -
kustomize build ./examples/kustomize/overlays/prod | k1c apply -f -

23 Cloudflare CRD kinds + 9 standard k8s kinds are wired today. "e2e" means the row is exercised by tests/e2e/idempotency.e2e.test.ts against a real Cloudflare account — round-trip create → read → equals → delete. "verified" rows were checked manually but aren't in the suite (they need a customer domain / paid feature).

See docs/resources.md for the full mapping and limitations, and TODO.md for what's queued.

drift-fixed = provider has a custom equals that strips Cloudflare-side defaults (e.g. proxied: false, storage_class: 'Standard', recording.mode: 'off') so re-applying an unchanged manifest stays NOOP. The pattern was uncovered while running the e2e suite against a real account — see PRs #25 / #27 / #29 / #35 / #36 / #37 / #38.

For users who want kubectl apply to actually reach Cloudflare (not just sit in etcd), k1c operator run watches CRD instances + label-gated standard resources and reconciles them via the same lower / plan / apply pipeline the CLI uses. See examples/k1c-operator/ for the install bundle (ServiceAccount + ClusterRole + Deployment), the helm chart, and an Argo CD GitOps install example.

GitOps (Argo CD / Flux) ──► etcd ──watch──► k1c operator ──► Cloudflare API

The operator runs the same code path as the CLI, so a feature added to either ships to both at once. Production-grade plumbing:

• Watch streams: subscribes to every Cloudflare CRD plural + label-gated standard kind, debounced 500ms; --interval doubles as a resync safety net; --no-watch falls back to pure polling.
• Status writeback: after each reconcile pass, patches .status.conditions on every touched Cloudflare CRD so kubectl get r2bucket reflects Reconciled / ReconcileFailed plus the underlying error message; .status.cloudflareNativeId is persisted for the cleanup path.
• Finalizer cascade delete: every CR carries a k1c.io/cleanup finalizer. kubectl delete flips deletionTimestamp; the operator deletes the corresponding Cloudflare resource and strips the finalizer; k8s GCs the CR. Failures retry on the next tick.
• Leader election (coordination.k8s.io/v1 Lease): scale replicas to 2+ for HA, leader/standby spread by topologySpreadConstraints. Failover within ~15s.
• Prometheus /metrics + /healthz + /readyz on :9090. ServiceMonitor template + Grafana dashboard included. Metrics: k1c_operator_{up, is_leader, reconcile_total, reconcile_passes_total, reconcile_duration_seconds, watch_events_total, managed_resources, finalizer_total}.
• Structured JSON logging via --log-format json for log aggregators.
• Graceful shutdown: SIGTERM waits up to 30s for the in-flight reconcile to finish before exiting.

PKL is a configuration language with strong types and module imports. PKL manifests catch typos / missing fields / out-of-range enums at edit time instead of at apply time:

// examples/pkl/hello-worker.pkl
import "../../pkl/k1c.pkl"

output {
  renderer = new YamlRenderer { isStream = true }
  value = resources
}

resources = new Listing {
  new k1c.R2Bucket {
    metadata { name = "media" }
    spec { location = "weur" }   // ← `"antarctica"` fails with line+col
  }
  new k1c.Deployment { ... }
}

# Either pipe pkl yourself …
pkl eval --format yaml examples/pkl/hello-worker.pkl | k1c apply -f -

# … or just hand the .pkl file directly. The CLI shells to
# `pkl eval --format yaml` when it sees a `.pkl` extension.
k1c apply -f examples/pkl/hello-worker.pkl

Three example layouts under examples/pkl/ show how far the typed approach scales:

CI runs k1c apply --validate-only on every .pkl example (k8s-validate.yml) so a rename in the zod schemas, a CSI driver change, or a PKL syntax slip fails the build before it lands. Reproduce locally with:

just validate-pkl

Hand-written modules ship under pkl/ and currently cover the most-used kinds (R2Bucket / KVNamespace / D1Database / Queue / Vectorize / ConfigMap / Secret / Deployment with CSI volumes). The long tail of CRDs (rulesets, Access*, etc.) stays YAML-only until k1c export-pkl can generate the full set from the zod schemas.

The CLI compiles to a WebAssembly Component via pnpm build:wasm. The pipeline:

1. tsc emits the JS sources.
2. esbuild bundles dist/cli/wasm-main.js (a Node-free entry that drops operator, logs, port-forward, rollout, config, and apply --watch — see src/cli/wasm-main.ts) into a single ~1.8 MB ESM module at dist-wasm/k1c.bundle.mjs.
3. componentize-js (optional) wraps the bundle in a wasi:cli/run component at dist-wasm/k1c.wasm. This step needs the wasi-cli WIT packages vendored under wit/deps/ — see the wit/README.md for the vendoring procedure. If the deps are missing the script logs the failure but still emits the bundle, so you can re-componentize out-of-band.

The bundle works in any wasi-cli host once wrapped. apply / diff / get / describe / delete / telemetry / explain / export-crds / version are the supported commands.

A k1c manifest is a subset of valid k8s YAML — the same file applies to either k1c or to a real kubectl cluster. Cloudflare-specific data sources ride on the standard volumes[].csi shape with k1c driver names (r2.k1c.io, kv.k1c.io, dispatch-namespace.k1c.io, ...) and Cloudflare CRDs live under cloudflare.k1c.io/v1alpha1.

Worker binding names are resolved from CSI volumeAttributes in this order: binding, legacy bindingName, then an upper-snake derivation from the volume name (r2-media -> R2_MEDIA). mountPath stays a Kubernetes mount path and is not used as the Worker binding name. A container's image is the Worker script entrypoint path; for manifests that need a Kubernetes-looking placeholder, cloudflare.com/source.<container-name> can override it.

Both entry points share one schema source (src/manifest/schemas.ts, zod):

                         src/manifest/schemas.ts (zod)
                                     ↑
              ┌──────────────────────┴──────────────────────┐
              │                                             │
    k1c apply -f m.yaml                          kubectl apply -f m.yaml
    (CLI parses zod directly)            (apiserver validates against
              │                            openAPIV3Schema, derived
              │                            from the same zod by
              │                            `k1c export-crds`)
              │                                             │
              │                                CR stored in etcd
              │                                             │
              │                            operator watches → reconciles
              │                            (re-parses with the same zod
              │                             at src/operator/source.ts)
              │                                             │
              └──────────►  Cloudflare API  ◄───────────────┘
                            (same lower / plan / apply path)

To register the CRDs once on the target cluster:

k1c export-crds | kubectl apply -f -
kubectl apply -f my-manifest.yaml --dry-run=server   # validates schema

The CSI drivers are not registered on the cluster, so a real Pod stays pending — but the manifest is schema-valid and admission controllers accept it. To make kubectl apply actually reach Cloudflare, install the operator (see Operator section).

So a schema-breaking change shows up on every PR through whichever path it affects, and a feature added to either entry point ships to both at once.

The CLI accepts standard k8s kinds (Deployment / Service / ConfigMap / Secret / ...) and translates them into Worker / binding equivalents. A real cluster handles those as native k8s resources, not as Cloudflare ones. So:

• For Cloudflare-only resources (R2Bucket, KVNamespace, D1Database, Queue, Vectorize, ...): apply either via k1c apply or via kubectl apply (with the operator installed). Both paths converge.
• For Deployment etc.: k1c apply lowers it to a Cloudflare Worker; kubectl apply creates a real k8s Deployment. Pick the path that matches what you actually want.

The operator only watches cloudflare.k1c.io/v1alpha1 CRDs plus label-gated standard kinds — see src/operator/source.ts for the gate predicate. That keeps "I just wanted a regular k8s Deployment" cases out of the Cloudflare reconcile loop.

examples/ carries one self-contained manifest per supported pattern (multi-container Pod, blueGreen / canary Rollout, CronJob, StatefulSet / Durable Object, Hyperdrive, Queue producer + consumer, AI + Vectorize RAG stack, multi-tenant SaaS, zone hardening combo, full-stack web app, etc.). Browse examples/README.md for the full index. Each file is exercised by src/examples-smoke.test.ts so the library stays in sync with the schemas.

I wanted a kubectl apply UX for personal Cloudflare projects but did not want to pay for a managed Kubernetes control plane (GKE minimum is roughly JPY 8,000 / month). k1c is the smallest tool that lets a Kubernetes-shaped manifest drive Cloudflare resources directly. See ADR-0001 for the full reasoning.

The architecture is documented as Architecture Decision Records (ADR-0001 through ADR-0007).

CLI (one-shot apply):

npm install -g @mizchi/k1c
# or:
pnpm dlx @mizchi/k1c apply -f manifest.yaml

Operator (continuous reconciliation inside a k8s cluster):

# Register CRDs once (kept out of the helm release lifecycle so
# `helm uninstall` doesn't orphan every CR you ever applied).
k1c export-crds | kubectl apply -f -

# API token Secret.
kubectl create namespace k1c-system
kubectl -n k1c-system create secret generic cloudflare-api-token \
  --from-literal=K1C_ACCOUNT_ID=<your-account-id> \
  --from-literal=CLOUDFLARE_API_TOKEN=<your-token>

# Either: flat install bundle.
kubectl apply -f https://raw.githubusercontent.com/mizchi/k1c/main/examples/k1c-operator/install.yaml

# Or: helm chart.
helm install k1c examples/k1c-operator/helm-chart \
  --namespace k1c-system --create-namespace=false

# Or: Argo CD Application — see examples/argocd/.

The operator image is published as a multi-arch (linux/amd64 + linux/arm64) OCI v1.1 image with SLSA provenance + SBOM attestations: ghcr.io/mizchi/k1c-operator:<version>.

From source (this repo):

pnpm install
pnpm test          # offline unit + integration tests (uses SDK mocks)
pnpm typecheck
pnpm build         # emits dist/

# Run the CLI in-repo (TypeScript via Node strip-types):
pnpm k1c apply   -f examples/hello-worker.yaml [--dry-run]

# Optional: end-to-end tests against a real Cloudflare account (creates and
# deletes resources). Requires K1C_E2E=1 plus the same env vars used by the CLI;
# the full suite expects the broad token recipe from docs/api-token.md.
K1C_E2E=1 K1C_ACCOUNT_ID=... CLOUDFLARE_API_TOKEN=... pnpm test:e2e

k1c apply    -f <file|dir|-> [--dry-run | --watch | --validate-only] [--quiet | -q]
k1c diff     -f <manifest.yaml> [-o text|json] [-v|--verbose] [--color always|never]
k1c delete   -f <manifest.yaml> [--cascade]
k1c get      <kind> [name] [-n <namespace>] [-o text|json]
k1c describe <kind> <name> [-n <namespace>] [-o text|json]
k1c rollout  {status|promote|abort} <ns>/<name> --dispatch <name>
k1c logs     <kind> <name> [-n <namespace>] [--format pretty|json] [--status <s>] [--limit N]
k1c port-forward <kind> <name> [-n <namespace>] [--port 8787]
k1c wrangler-config -f <manifest.yaml> [--worker <namespace/name>]
k1c telemetry workers <kind> <name> [-n <ns>] [--since 1h] [-o text|json]
k1c explain  <kind | list> [--recursive | -r]
k1c export-crds [--include-standard]
k1c config   list | current-context | use-context <name>
             | set-context <name> --account <id> [--zone <id>] [--token-env <env>]
             | delete-context <name>
k1c operator run [-n <namespace>] [--interval 30] [--no-watch]
                 [--leader-election] [--lease-name k1c-operator]
                 [--lease-namespace k1c-system]
                 [--metrics-addr 0.0.0.0:9090 | --no-metrics]
                 [--log-format text|json]
k1c version

logs and port-forward shell out to a locally-installed wrangler (wrangler tail and wrangler dev --remote, respectively). Resource kind must lower to a Worker (Deployment, Rollout, CronJob, Job, StatefulSet, or Worker itself).

wrangler-config is an offline bridge for local Worker dev. It parses and lowers the same manifest as apply, selects one lowered Worker, and prints a wrangler.jsonc-compatible JSON config with main, compatibility settings, and Worker bindings, including Workers AI ai, Dynamic Workers worker_loaders, Workers for Platforms dispatch_namespaces, and Agents Durable Object bindings/migrations. If the manifest lowers to multiple Workers, pass --worker default/api (or another <namespace>/<name> label). JSON is valid JSONC, so the output can be used directly as a Wrangler config. Resource IDs that k1c resolves only at apply time are omitted from the generated config, leaving binding names explicit while keeping local dev usable.

Authentication is via two environment variables:

For multi-account / multi-zone setups, use k1c config set-context to record each environment's account / zone / API token env-var name in ~/.k1c/config.yaml, then switch with k1c config use-context <name> or per invocation with --context <name>. Tokens themselves stay in env vars (apiTokenEnv only stores the var name) so the file is safe to share across machines.

manifest.yaml
   │
   ▼  src/manifest/parse.ts        — YAML → typed K1cResource[] (validated by zod)
   │
   ▼  src/manifest/lower.ts        — resolves refs, folds ConfigMap/Secret into Workers,
   │                                  generates DispatcherWorker / state KV for canary Rollouts
   ▼  src/reconciler/plan.ts       — compares desired vs actual via providers, topological sort
   │
   ▼  src/reconciler/apply.ts      — executes operations (create/update/delete) with retry
   │
   ▼  src/canary/runtime.ts        — for canary Rollouts: read KV state, run state machine,
   │                                  upload canary + rewrite weight + promote
   ▼  Cloudflare account

Providers live under src/providers/ and are uniform across resource types. The interface mirrors AWS CloudControl (CRUD + Status + List + Discovery); see ADR-0006.

This is experimental. In particular:

• Worker entrypoint content is hashed at lower time and round-tripped via the k1c.io/content-hash= script tag, so editing only the JS file (without changing the manifest) now triggers a Worker update on apply.
• Async polling: providers can return kind: 'async' from create / update; the apply loop polls status() until success / failure (used by CustomHostname).
• The reconciler model assumes a single Cloudflare account at a time.
• End-to-end coverage is opt-in (tests/e2e/**, env-gated by K1C_E2E=1). The default pnpm test exercises providers against SDK mocks only; running the e2e suite requires a real Cloudflare account and creates / deletes resources there. See Releases and Limitations below for what is and is not covered.

Versioning is automated via release-please:

• Conventional Commit messages on main (feat:, fix:, chore:, etc.) feed into a release PR that bumps package.json, updates CHANGELOG.md, and cuts a Git tag plus a GitHub Release.
• The same release-please.yml workflow then runs a publish job (gated on release_created == true) that publishes @mizchi/k1c to npm via OIDC trusted publishing (no NPM_TOKEN) with --provenance SLSA attestation.

The npm package must be registered as a trusted publisher on npmjs.com once, pointing at mizchi/k1c + the release-please.yml workflow. After that the entire flow (PR → merge → tag → npm) is hands-off.

MIT