crewkit

Framework de context engineering para desenvolvimento assistido por IA. Um comando para configurar agents, skills, hooks, rules e memory no seu projeto.

O que faz

O crewkit escaneia sua codebase e gera um setup completo e calibrado de context engineering:

• Agents (5) — explorer, architect, coder, tester, reviewer com roles definidos e tiers de modelo
• Skills (4) — full-workflow, hotfix, explore-and-plan, review-pr
• Hooks (4) — session-start, protect-sensitive-files, post-compact-recovery, stop-quality-gate
• Rules — regras de codigo por stack geradas a partir de padroes detectados
• Memory (.ai/memory/) — architecture, conventions, commands, testing, lessons
• CLAUDE.md — hard rules do projeto derivadas de padroes reais do codigo
• settings.json — permissoes conservadoras + registro de hooks
• .mcp.json — MCP servers detectados da sua infraestrutura (PostgreSQL, Redis, Sentry, etc.)

Tudo calibrado pro seu codigo real — nao e boilerplate generico.

Instalacao

npx crewkit install

Copia a skill para ~/.claude/skills/crewkit-setup/ (setup unico).

Uso

Abra qualquer projeto no Claude Code e rode:

/crewkit-setup

A IA escaneia sua codebase (~8-15 min) e gera o setup completo. Zero perguntas durante o scan — detecta tudo a partir do codigo, historico git e docs.

O que acontece

1. Pre-flight — verifica modelo, detecta re-runs, oferece backup
2. Phase 0 — cria estrutura de diretorios, atualiza .gitignore
3. Phase 1-3 — detecta stacks, le docs, mapeia estrutura do projeto
4. Phase 4 — leitura profunda de arquivos representativos, detecta padroes com score de confianca
5. Phase 5 — health check (opcional, com --check-health)
6. Phase 6 — apresenta perfil pra sua revisao antes de gerar
7. Phase 7 — gera todos os arquivos em ordem de falha graceful

Re-runs

Se voce rodar /crewkit-setup de novo num projeto que ja tem setup:

• [R] Regenerate — backup do existente, regenera tudo
• [M] Memory only — re-scan e atualiza memory + CLAUDE.md, preserva agents/hooks/skills
• [C] Cancel

Stacks suportadas

Detectadas automaticamente a partir de arquivos de manifesto:

| Stack | Detectado por | |-------|--------------| | .NET | *.csproj, *.sln | | Node.js | package.json (Express, Next.js, Nest, etc.) | | Python | pyproject.toml, requirements.txt | | Go | go.mod | | Rust | Cargo.toml | | Java/Kotlin | pom.xml, build.gradle | | Ruby | Gemfile | | Elixir | mix.exs | | PHP | composer.json | | Dart/Flutter | pubspec.yaml |

Agents gerados

| Agent | Modelo | Funcao | |-------|--------|--------| | explorer | Sonnet | Exploracao read-only da codebase | | architect | Opus | Review de arquitetura + gate de aprovacao | | coder | Sonnet | Diffs minimos, padroes existentes | | tester | Sonnet | Workflow completo de testes + accountability de cobertura | | reviewer | Opus | Code review de alto sinal, sem ruido |

Skills opcionais

Alem das 4 skills core (instaladas automaticamente), voce pode adicionar skills opcionais:

npx crewkit add retro          # post-mortem de tarefas
npx crewkit add dev-metrics    # metricas de desenvolvimento via git
npx crewkit add security-scan  # scan OWASP top 10
npx crewkit add impact         # analise de blast radius

IDEs suportadas

• Claude Code — suporte completo (agents, skills, hooks, rules, memory)
• GitHub Copilot — auto-detectado, gera copilot-instructions.md, agents e prompts
• Cursor — auto-detectado, gera rules com globs e AGENTS.md

A IDE e detectada automaticamente durante o setup. .ai/memory/ e compartilhado entre todas as IDEs.

O que e commitado vs ignorado

Commitado (compartilhado com o time):

• CLAUDE.md, .claude/agents/, .claude/rules/, .claude/skills/, .claude/hooks/
• .claude/settings.json, .ai/memory/

Ignorado (pessoal/metadata):

• .claude/settings.local.json, .crewkit/, .crewkit-backup/

Requisitos

• Node.js >= 20
• Claude Code, GitHub Copilot ou Cursor
• Opus ou Sonnet recomendado (Haiku funciona mas gera scans mais rasos)

Licenca

MIT

English

Context engineering framework for AI-assisted development. One command to set up agents, skills, hooks, rules, and memory for your project.

What it does

crewkit scans your codebase and generates a complete, calibrated context engineering setup:

• Agents (5) — explorer, architect, coder, tester, reviewer with defined roles and model tiers
• Skills (4) — full-workflow, hotfix, explore-and-plan, review-pr
• Hooks (4) — session-start, protect-sensitive-files, post-compact-recovery, stop-quality-gate
• Rules — per-stack coding rules generated from detected patterns
• Memory (.ai/memory/) — architecture, conventions, commands, testing, lessons
• CLAUDE.md — project-specific hard rules derived from actual code patterns
• settings.json — conservative permissions + hook registrations
• .mcp.json — MCP servers detected from your infrastructure (PostgreSQL, Redis, Sentry, etc.)

Everything is tailored to your actual code — not generic boilerplate.

Install

npx crewkit install

This copies the skill to ~/.claude/skills/crewkit-setup/ (one-time setup).

Usage

Open any project in Claude Code and run:

/crewkit-setup

The AI scans your codebase (~8-15 min) and generates the full setup. No questions asked during scan — it detects everything from code, git history, and docs.

What happens

1. Pre-flight — checks model, detects re-runs, offers backup
2. Phase 0 — creates directory structure, updates .gitignore
3. Phase 1-3 — detects stacks, reads docs, maps project structure
4. Phase 4 — deep reads representative files, detects patterns with confidence scoring
5. Phase 5 — health check (optional, with --check-health)
6. Phase 6 — presents profile for your review before generating
7. Phase 7 — generates all files in graceful-failure order

Re-runs

If you run /crewkit-setup again on a project that already has a setup:

• [R] Regenerate — backup existing, regenerate everything
• [M] Memory only — re-scan and update memory + CLAUDE.md, preserve agents/hooks/skills
• [C] Cancel

Supported stacks

Detected automatically from manifest files:

| Stack | Detected by | |-------|------------| | .NET | *.csproj, *.sln | | Node.js | package.json (Express, Next.js, Nest, etc.) | | Python | pyproject.toml, requirements.txt | | Go | go.mod | | Rust | Cargo.toml | | Java/Kotlin | pom.xml, build.gradle | | Ruby | Gemfile | | Elixir | mix.exs | | PHP | composer.json | | Dart/Flutter | pubspec.yaml |

Generated agents

| Agent | Model | Role | |-------|-------|------| | explorer | Sonnet | Read-only codebase exploration | | architect | Opus | Architecture review + approval gate | | coder | Sonnet | Minimal diffs, existing patterns | | tester | Sonnet | Full test workflow + coverage accountability | | reviewer | Opus | High signal code review, no noise |

Optional skills

In addition to the 4 core skills (auto-installed), you can add optional skills:

npx crewkit add retro          # task post-mortem
npx crewkit add dev-metrics    # development metrics from git
npx crewkit add security-scan  # OWASP top 10 scan
npx crewkit add impact         # blast radius analysis

Supported IDEs

• Claude Code — full support (agents, skills, hooks, rules, memory)
• GitHub Copilot — auto-detected, generates copilot-instructions.md, agents and prompts
• Cursor — auto-detected, generates rules with globs and AGENTS.md

IDE is auto-detected during setup. .ai/memory/ is shared across all IDEs.

What gets committed vs ignored

Committed (shared with team):

• CLAUDE.md, .claude/agents/, .claude/rules/, .claude/skills/, .claude/hooks/
• .claude/settings.json, .ai/memory/

Ignored (personal/metadata):

• .claude/settings.local.json, .crewkit/, .crewkit-backup/

Requirements

• Node.js >= 20
• Claude Code, GitHub Copilot, or Cursor
• Opus or Sonnet recommended (Haiku works but produces shallower scans)

License

MIT