Ray Plugins — Marketplace de Exemplo Claude Code
RayDataCo/ray-plugins
Template educacional para marketplaces de plugins Claude Code
Ver no GitHubTemplate mínimo e funcional para criar um marketplace de plugins Claude Code. Inclui exemplos de skills básicas (/hello, greeting) e estrutura completa para integração com o ecossistema Ray Data Co, servindo como ponto de partida para desenvolvedores que querem construir plugins próprios.
Instalação
npx convez-ai RayDataCo/ray-plugins1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Aprender a estrutura de um plugin Claude Code
- Criar um marketplace de plugins corporativo
- Desenvolver skills reutilizáveis em projetos Ray Data
Veredito

Repositório útil como referência educacional para entender a arquitetura de marketplaces Claude Code, mas com conteúdo limitado em produção real. Melhor aproveitado por desenvolvedores que estudam a estrutura antes de construir plugins complexos.
README
A tradução para o português deste README ainda está na fila.
ray-plugins — Ray Data Co's example Claude Code plugin marketplace
A minimal, working example of a Claude Code plugin marketplace. Add it and install the example plugin in two commands.
Use it
/plugin marketplace add RayDataCo/ray-plugins
/plugin install rdco-example@ray-plugins
Then say hello to it, or ask for a Ray Data Co greeting — both are skills.
What's inside
.claude-plugin/marketplace.json # marketplace manifest (lists plugins)
plugins/rdco-example/ # minimal copy-me template (skills-only)
.claude-plugin/plugin.json # plugin manifest
skills/rdco-greeting/SKILL.md # a skill (explains plugin structure)
skills/rdco-hello/SKILL.md # a skill (say hello, by name)
plugins/copilot-studio-agent-builder/ # real plugin: build Microsoft Copilot Studio agents (skills-only)
.claude-plugin/plugin.json
skills/copilot-studio-agent-builder/ # lifecycle orchestrator (plan->build->test->publish->maintain)
SKILL.md
reference/ # generic runbook (living) + verified facts + doc-endpoint index
templates/ # the 4 per-agent deliverable templates (#2-#5)
skills/copilot-studio-pull-runbook/ # skill: pull the runbook into the current project
skills/copilot-studio-ground-runbook/ # skill: re-ground the runbook in the latest MS docs
plugins/ab-skill-factory/ # real plugin: the agent brigade that manufactures skills
.claude-plugin/plugin.json
README.md + DESIGN.md + BUNDLE-SPEC.md + RAIL-SPEC.md # pattern docs + interface specs
skills/expo/ # the deciding role at the pass (routing + exit set)
skills/execution-eval-station/ # lift-over-base-model value gate (per-fixture, per-tier)
workflow/*.run.js # reference workflow runs (as-executed)
examples/ # worked examples w/ real run reports (variance-analysis, generate-tests)
plugins/ab-managerial-accounting/ # discipline brigade: mise + expo + service over finance stations
.claude-plugin/plugin.json
MENU.md # the station roster (situation -> station / base-model-covered)
skills/mise/ # readiness gate (mise.py engine + mise.toml)
skills/expo/ # the composing coordinator (decompose -> select -> compose)
skills/service/ # the mise-gated on/off switch
skills/variance-analysis/ # + 4 more eval-proven finance stations
base-model-covered/ # tasks the base model covers (exemplar prompts + evidence)
plugins/ab-data-engineering/ # discipline brigade: mise + expo + service over DE stations
.claude-plugin/plugin.json
MENU.md
skills/mise/ + skills/expo/ + skills/service/ # the standard brigade surface
skills/pipeline-failure-triage/ # first eval-proven DE station
plugins/ab-legal/ # discipline brigade: structural (never advisory) legal review
MENU.md + skills/mise|expo|service/ # the standard brigade surface
skills/regulatory-notice-coverage/ + skills/contract-clause-coverage/ # eval-proven stations
evals/ # two-arm execution-eval evidence
plugins/ab-domain-research/ # FILL brigade: stocks a cellar with domain competency + gold exemplars
MENU.md + skills/mise|expo|service/ # the standard surface (mise incl. CELLAR_ROOT check)
skills/cert-body-sourcing/ + standards-regulatory-sourcing/ + academic-ocw-sourcing/ + public-filings-exemplars/ + award-case-study-exemplars/
evals/ # two-arm execution-eval evidence (both rounds)
plugins/ab-marketing/ # discipline brigade: first exemplar-graded generative station
MENU.md + skills/mise|expo|service/ + skills/effectiveness-narrative-draft/ + evals/
Plugins
rdco-example— minimal template: two example skills (rdco-greeting+rdco-hello), skills-only. Copy it to start your own.copilot-studio-agent-builder— guides the full lifecycle of building a Microsoft Copilot Studio agent and emits five standard deliverables (generic runbook + per-agent implementation / evaluation / documentation / maintenance docs). Generative-orchestration baseline; Work-IQ-off + grounded-internal defaults; built from a verified Microsoft Learn walkthrough. Works in Claude Code (deliverables as markdown files) and Claude Web/Desktop (deliverables as project assets). Skills-only (no slash commands) — ships two companion skills:copilot-studio-pull-runbook(pull the runbook into your project) andcopilot-studio-ground-runbook(refresh it against the latest Microsoft docs — the field moves fast)./plugin install copilot-studio-agent-builder@ray-pluginsab-skill-factory— the agent brigade that manufactures Claude skills: stations (spec → tests → author → critic) coordinated at the pass by an expo that decides routing via a closed exit set (advance · refire-to-author · reroute-to-spec · kill), pulling mutable tickets off a rail. Includes theexecution-eval-station(measures a produced skill's lift over the base model, per-fixture and per-tier), a deterministic skill-lint gate, full pattern docs (README/DESIGN/BUNDLE-SPEC/RAIL-SPEC), reference workflow runs, and two worked examples with real measured-lift reports.ab-registrar— the first house role (a coordinator/record-keeper, not a kitchen brigade): shipsmise+steward(intake, context curation, close-out sweep — moved here from the factory),registry(menu catalog), andorders(open-orders report). Noservice/fire/stations — a house role never works a ticket.ab-managerial-accounting— managerial-accounting discipline brigade (nédiscipline-skills): the standard brigade surface —mise(readiness gate) +expo(the composing coordinator that decomposes a request, fires the finance stations it needs, and synthesizes one answer, incl. compound asks like a full due-diligence picture) +service(mise-gated on/off) — over 5 eval-proven finance stations (variance-analysis,annual-budget-build,close-management,treasury-liquidity-analysis,debt-schedule), plusMENU.md(the station roster) and thebase-model-covered/registry (5 tasks with eval-verified base-model coverage). Only eval-passers ship.ab-data-engineering— data-engineering discipline brigade: the samemise+expo+servicesurface over eval-proven DE stations. First stationpipeline-failure-triage(+1.00 lift on silent zero-row failure classification);MENU.mdcarries honest per-task status for the rest (held-for-refire / weak-evidence / base-model-covered-pending). Only eval-passers ship.ab-legal— legal-review discipline brigade, structural not advisory: themise+expo+servicesurface over eval-proven document-checking stations. Two live —regulatory-notice-coverage(GDPR/CCPA disclosure-element coverage) andcontract-clause-coverage(standard clause-set coverage by contract type, +1.00 lift) — plusredline-against-playbookheld-for-refire. Every station checks a document against an explicit enumerated standard and reports gaps; it never gives legal advice.ab-domain-research— the fill brigade (kitchen kind, domain-centric analog of company research): stocks a knowledge cellar with domain competency and gold exemplars — the two things a generative-skill build needs, since generative skills have no computable oracle and get graded against exemplars. Standard surface (miseincl. aCELLAR_ROOTlanding-target check, composingexpo,service) over 6 eval-proven stations:cert-body-sourcing(+0.60),standards-regulatory-sourcing(+1.00),academic-ocw-sourcing(+0.60, CC-licensed-only gate),public-filings-exemplars(+0.40–0.60 — EDGAR exhibits/MD&A, SAM.gov RFP/SOWs, court briefs, patents, GAO/CRS, each exemplar with awhy_gold),award-case-study-exemplars(+0.60 — creative-domain gold exemplars from juried award libraries), andfreshness-watch(99% per-element — re-verifies landed grounding against PRIMARY sources on a cadence and routes factory refire tickets: EVAL-SPEC's world-change regression trigger, live). Every station enforces license-gate-before-content, authority tiering, per-source dispositions with reason classes, provenance frontmatter, and untrusted-data discipline (embedded instructions ⇒injection-suspect). Curation, never bulk scraping; eval evidence inevals/.ab-marketing— marketing discipline brigade: themise+expo+servicesurface over eval-proven marketing stations. Three live generative stations — the marketplace's first exemplar-graded generative skills (eval rubrics derived from real GAO campaign-evaluation + S-1 positioning gold sourced by ab-domain-research the same morning):effectiveness-narrative-draft(+1.00),positioning-narrative-draft(+0.80/+1.00 across rounds; round-1's critic-caught license violation fixed and re-verified),marketing-brief-draft(+0.80 both rounds). Plusadvertising-claims-endorsement-review(structural FTC 255 coverage review, base 33%→93% per-element under the founder-ruled measure-skill-quality bar).MENU.mdholds an honest 4-round evidence trail for the 2 still-held structural stations. Structural stations report elements, never legal advice. Only eval-passers ship.
Copy it
plugins/rdco-example/ is a template. Copy the folder, rename it, edit plugin.json, add your own skills/, agents/, hooks, or MCP servers, then add an entry to .claude-plugin/marketplace.json. (RDCO standardizes on skills over slash commands — build capabilities as skills.)
Other sources
- Private GitHub repos work — Claude Code uses your existing
gh/git credentials (anyone with repo access can add it). For background auto-updates, setGITHUB_TOKEN. - Bitbucket / GitLab / self-hosted work via a full git-URL source, e.g.
/plugin marketplace add https://bitbucket.org/team/plugins.git.
Built by Ray Data Co.