Skip to content

Client value

Fewer decisions to remember, fewer places for mistakes to hide, and every outward step reviewable.

  • Time to first deploy: one slash command creates a repository that already builds, tests, scans and publishes a multi-arch image; the GitOps repository runs it locally in a few minutes (devbox run cluster-up took about 1m40s from no cluster to every ArgoCD application healthy in the platform’s own test).
  • You do not become a platform team. Secrets, a database, telemetry and policy are declarations in one file, not projects.
  • Reviewable change: adding a service or promoting a version is a pull request with the generated diff, not a sequence of commands you hope you remembered.
  • Leave any time: the output is plain Kubernetes, Kustomize, ArgoCD, Gateway API and OpenTelemetry; the platform is Apache-2.0.
  • /component:spec → /component:plan → /component:build (spec-driven pipeline) and /component:fix (small change) work the same in Python, Java, Go and Next.js repositories; the right agents are routed automatically.
  • You never need to know the manifests. A service repository contains code and a Dockerfile, nothing about Kubernetes.
  • devbox run <recipe> runs identically on your laptop, in CI and for agents.
  • Everything that can be code is tested code. The cplat CLI has an automated test suite, and an end-to-end test builds three repositories, a cluster and the whole flow in CI.
  • Defaults are policy. The conventions (non-root, read-only filesystem, limits, allowed registries, no latest outside dev) are checked in the pull request and enforced in the cluster.
  • Extensible by contract: a new language is a new shape (a registry entry, a template, a plugin) following a documented contract; a new resource type implements a stable connection contract, so your services never depend on the implementation.
Control What the platform does Evidence
Secrets Created in the cluster by External Secrets Operator; never stored in git ADR-018
Supply chain CI actions pinned by commit SHA, least-privilege tokens, Trivy image scan that blocks release builds, CycloneDX SBOM per architecture ADR-019
Runtime hardening Numeric non-root user, read-only filesystem, no privilege escalation, capabilities dropped Generated manifests, ADR-017
Guard rails Kyverno policies evaluated offline in your pull request and enforced in the cluster ADR-022
Change control Promotion and composition happen through reviewed pull requests; a protected main Promote safely
Licenses REUSE-compliant repository, third-party notices, SBOMs for generated images Third-party notices

This supports an audit; it is not a certification. You remain responsible for your own controls.

  • CI minutes: pull requests run only what a change needs; documentation-only changes run the cheap validation jobs, superseded runs are cancelled and dependency updates arrive monthly in groups.
  • Agent cost: always-on prompt text shrank by 28 percent and the product-manager agent runs on a smaller model; scripts replace long prompts, so common commands spend fewer tokens. Headless runs start on a Sonnet coordinator: a small task measured about $0.48 before and $0.15 after, a full build $1.94 and $1.26, and a full ship $1.20 to $1.60 (first measured, LAN cluster); one budget per run ($15 per ship) caps the worst case (estimated list prices, one benchmark repository).
  • No platform to license or host: it is a repository, a CLI and plugins.
Claim Number Source
Defects found by running the real chain, each now guarded 12 Changelog 1.1.1
Multi-arch Next.js image build about 9 min to about 3 min Baselines
Multi-arch Java image build about 3 min to about 2 min Baselines
Always-on prompt tokens 28 percent fewer Changelog
Local cluster from nothing to all applications healthy about 1m40s Measured on the test application

Measured on one test application and Apple-silicon laptops with GitHub-hosted runners; your numbers will differ, the method is in the baselines page.

  • Claude Code and GitHub are assumed. The commands are Claude Code plugins, the templates target GitHub Actions and GHCR. Other hosts need work.
  • Slash commands are prompts around scripts. The scripts are tested; the prompts that call them are not deterministic and always ask before outward actions.
  • Cross-repository work is planned centrally, built per repository. /system:spec and /system:plan produce a system spec and a validated plan that assigns every criterion to a repository; each repository builds its part from its own spec, and nothing merges without you.
  • The cluster is yours. The platform installs operators and ArgoCD on a local k3d cluster; real clusters are bootstrapped by a human once (devbox run bootstrap), and operators (External Secrets, CloudNativePG, Kyverno) must exist there if you use their features.
  • Resources are deliberately small. The Postgres resource has no backups, point-in-time recovery or pooling; observability collects and forwards but ships no alert rules; Go services expose Prometheus metrics but no OpenTelemetry traces yet.
  • Guard rails cover what the platform renders. Policies check Deployments; pods created by operators are out of scope.
  • Single maintainer. It is an open-source project with best-effort support.

Next: usage scenarios or get started.