Claude Code Agent Builder
ZSlayerHQ/claude-code-agent-builder
Scaffolder de projetos multi-agente production-grade para Claude Code
Ver no GitHubMeta-agente que scaffolda projetos em produção para Claude Code, gerando automaticamente estruturas disciplinadas com CLAUDE.md tuneado, subagentes especializados, hooks, configuração MCP e templates de documentação. Faz intake de 4-5 perguntas, propõe um roster de agentes e gera a árvore de diretórios completa e funcional.
Instalação
npx convez-ai ZSlayerHQ/claude-code-agent-builder1 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
- Estruturar projetos complexos com múltiplos subagentes especializados
- Implementar padrões de segurança e tool-scoping em sistemas de IA
- Automatizar setup de ambientes com MCP servers e plugins pré-configurados
- Estabelecer governança e auditoria em projetos baseados em agentes
Veredito

Ferramenta sofisticada para times que precisam orquestrar múltiplos agentes Claude com disciplina production-grade. Excelente para reduzir overhead de coordenação e implementar tool-scoping como defesa. Curva de aprendizado íngreme, mas ROI significativo em projetos complexos.
README
A tradução para o português deste README ainda está na fila.
Claude Code Agent Builder
A meta-agent for Claude Code that scaffolds disciplined, production-grade project directories — complete with a tuned CLAUDE.md, specialist subagents, settings, hooks, MCP config, and session-doc templates.
You point it at a new project idea; it asks 4-5 questions, proposes an agent roster, and generates the whole working tree.
You: "I need a Claude Code setup for a personal-finance tracker."
Bot: [4-5 intake questions]
Bot: [Proposes 3-4 agents with role + tools]
You: "Approved."
Bot: [Generates the full directory into output/]
What you get
Each generated project ships with:
CLAUDE.md— domain-aware identity, operational principles, agent roster, file locations (~80-150 lines).claude/agents/*.md— 3-5 specialist subagents, each with the right tool scope (Builders write, Auditors don't, etc.).claude/settings.json— Opus 4.8 + xhigh effort + 6 default plugins (superpowers,feature-dev,code-review,frontend-design,context-mode,skill-creator) + deny list + PreCompact + PostToolUse hooks.mcp.json— 4 default MCP servers (context7,gitnexus,playwright,scraplingviauvx).claude/hooks/—pre-compact-state-write.mjs(auto-updates STATE.md before context compaction) +prompt-injection-scan.mjs(scans tool-response inputs for known injection patterns).claude/skills/— operator-facing README + starterSKILL.mdtemplatesession-docs/—STATE.md/SESSION-LOG.md/GOTCHAS.md/DECISIONS.md/AUDIT-CHECKLIST.mddocs/tool-sources.md— install URLs + activation notes for every default tooldocs/research/— wave-based research index, populated if you opted into a research wave at intakestart.bat— one-click session launcher with-n "<project-name>"so/resumestays readableinit.sh— operator-customisable project setup script- Optional: lexicon kit (19-layer schema-of-schemas + drift register + naming rule) for app-shaped projects with ≥5 domain entities
- Optional:
docs/VISION.mdfor projects with mission scope bigger than v1
Design principles
- Four-agent ceiling. Most projects need 3-5 agents. Anything beyond 7 needs explicit justification — benefits plateau at 4 and ~80% of multi-agent failures come from coordination overhead, not missing capability.
- Tool scoping is the primary safety value. Auditors and Reviewers never get
WriteorEdit. Builders get a curated set. Researchers get web + Context7. The roster is shaped by what each agent can touch, not just what it knows. - Positive delegation triggers. Per Anthropic's Opus 4.8 guidelines: agent files say "Delegate to X when …" not "Avoid over-delegating." Opus 4.8 under-spawns by default; positive framing calibrates it.
- Research-wave dispatch. When a new project has 3+ orthogonal research territories (third-party APIs, regulatory boundaries, competitor sweeps, library evals, domain rule sets), the builder dispatches a parallel subagent wave — one subagent per territory — and synthesises into
docs/research/00-recommendations.mdbefore proposing the roster. - Lexicon discipline. App-shaped projects get a
lexicon/resources/_schema.yml+ an auto-loaded naming rule. Code generates from the lexicon; the lexicon does not generate from code. - Tight outputs. No AUP-redundant policy padding (Anthropic's system prompt already handles it). No no-op deny rules (audit your MCP tool surface first). See
references/claude-md-slimming-guide.md.
Quick start
You'll need Claude Code installed. The builder runs as a Claude Code session inside this repository.
# Clone
git clone https://github.com/ZSlayerHQ/claude-code-agent-builder.git
cd claude-code-agent-builder
# Open in Claude Code
claude --dangerously-skip-permissions -n "agent-builder"
Then in the session:
"Build me a Claude Code setup for a {your project description}"
The builder will run the 4-5 intake questions, propose a roster, and on approval generate the full project tree into output/<project-name>/. From there you typically cp -r output/<project-name> ../<project-name>/ to promote it to a sibling directory and start working in the new project.
What this repo is
CLAUDE.md— the builder's own identity + operating rulestemplates/— every starting point the builder copies into a generated projectreferences/— the builder's knowledge base (15 reference docs covering Anthropic guidelines, decomposition heuristics, tool scoping, MCP design, security patterns, the slimming guide, research-wave dispatch, lexicon discipline, etc.)scripts/check-template-paths.sh— drift guard. Run before committing template changes — catches agent-builder-internal paths that would dangle in every generated project..claude/rules/— auto-loaded operating rules (identity + output-format)session-docs/—GOTCHAS.md(8 real failure modes with fixes) +AUDIT-CHECKLIST.md(8-step honest-audit procedure)
Modes
The builder has three operating modes:
| Mode | Trigger | Output |
|---|---|---|
| Full-Suite | "Build me a setup for X" / "New project: X" | Complete directory generated into output/<name>/ |
| Individual | "Add a {researcher / auditor / reviewer} agent to {existing project}" | One agent file added to the existing roster |
| Audit | "Is it ready?" / "Audit this" / "Does X look solid?" | Runs AUDIT-CHECKLIST.md end-to-end against the artifact |
See CLAUDE.md for the full flow of each mode.
Versioning
This repo follows the Claude model generation it targets. Current target: Opus 4.8 / Sonnet 4.6 / Haiku 4.5 (current generation). When Anthropic ships a new generation, the relevant reference docs are updated and the templates pinned to the new model IDs.
Contributing
See CONTRIBUTING.md for the contribution flow + scope guidance. Issues and PRs welcome — particularly:
- New agent archetype templates
- Domain-specific reference docs (compliance scaffolds, niche stacks)
- Additional default MCP server configurations
- Refinements to the research-wave dispatch pattern
- Documentation improvements
License
Apache-2.0 — permissive, with patent grant. See NOTICE for third-party attribution.