Client value
Fewer decisions to remember, fewer places for mistakes to hide, and every outward step reviewable.
By role
Section titled “By role”Founder or lead of a small team
Section titled “Founder or lead of a small team”- 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-uptook 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.
Developer working on one service
Section titled “Developer working on one service”/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.
Platform or DevOps engineer
Section titled “Platform or DevOps engineer”- Everything that can be code is tested code. The
cplatCLI 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
latestoutside 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.
Security and compliance
Section titled “Security and compliance”| 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.
Finance and operations
Section titled “Finance and operations”- 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.
Evidence
Section titled “Evidence”| 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.
Honest limits
Section titled “Honest limits”- 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:specand/system:planproduce 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.