convez.ai
AgenteAutomaçãoIntermediário
fivetaku

kkirikkiri — Construtor de Times IA com Linguagem Natural

fivetaku/kkirikkiri

Uma frase. Um time de agentes montado e executando.

Ver no GitHub
InicialPluginScore · de 0 a 100

Agent que constrói times de agentes Claude automaticamente a partir de uma descrição em linguagem natural. Realiza entrevista rápida (2-3 perguntas), escaneia o ambiente local, propõe a equipe ideal e executa — tudo dentro do Claude Code com validação e memória compartilhada.

Instalação

$npx convez-ai fivetaku/kkirikkiri

1 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

Estrelas
63
Forks
12
Acompanhando
0
Saúde
28%
Último commit
há 1 semana
Versão
v0.26.5
Issues abertas
0
Contribuidores
1

Casos de uso

  • Montar times de pesquisa para análise de dados complexos
  • Criar equipes de desenvolvimento para implementar features multi-agente
  • Organizar times de análise e auditoria de código ou processos
  • Construir grupos de redação e revisão para documentação/conteúdo
  • Assemblar times de PM/estratégia para planejamento de produto

Veredito

Ferramenta inovadora que elimina configuração manual de multi-agentes via YAML. Excelente para prototipagem rápida de times IA, mas requer habilitação experimental de Agent Teams no Claude Code e ecossistema específico (Codex, Antigravity CLI) para máximo potencial.

README

A tradução para o português deste README ainda está na fila.

English | 한국어

kkirikkiri (끼리끼리)

One sentence. A team of AI agents, assembled and running.

Describe what you want in plain language. kkirikkiri interviews you with 2–3 questions, scans your environment, proposes a team, and executes — all within Claude Code.

Quick Start • Why kkirikkiri? • How it works • Features • Requirements


Quick Start

1. Add the marketplace

/plugin marketplace add https://github.com/fivetaku/gptaku_plugins.git

2. Install

/plugin install kkirikkiri

3. Enable Agent Teams

// ~/.claude/settings.json
{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}

4. Run

/kkirikkiri build me a research team

Why kkirikkiri?

  • Natural language in, running team out — no YAML, no agent definitions to write by hand
  • Interview-driven — 2–3 targeted questions replace a long configuration form
  • Environment-aware — detects installed tools (Codex CLI, Antigravity CLI agy, .claude/agents/) and builds the best team from what you actually have
  • Multi-model — Claude, Codex CLI (code & large-scale analysis), and Antigravity CLI (design/UI) can each take different roles in the same team
  • Two execution substrates — you pick: a live collaborating team (Agent Teams) or a deterministic agent pipeline (Workflows) for high-volume fan-out work
  • Validation loop — if round 1 output falls short, the team is automatically retried or rebuilt (up to 3 rounds)
  • Shared memory — .kkirikkiri/teams/{team_name}/ files persist across rounds so a replacement team picks up context immediately; each session gets its own directory to prevent multi-session collisions
  • Reusable agents — save team members to .claude/agents/ for use in future projects

The name comes from the Korean idiom 끼리끼리 — like-minded people naturally gathering together. Every team is assembled around a shared purpose.


How it works

Natural language input
    → Step 1: Intent detection + preset matching
    → Step 2: Environment scan (parallel)
    → Step 3: Interview — 2–3 AskUserQuestion prompts
    → Step 4: Dynamic team composition
    → Step 5: Team proposal + your confirmation
    → Step 6: Shared memory init + team execution
    → Step 7: Quality validation loop (up to 3 rounds)
    → Step 8: Result collection + report

Team leader rules:

  • Leader is always the most capable model available (Opus by default)
  • Leader plans, delegates, and validates — never writes code directly
  • Each member has a strictly scoped role

Features

Presets

Five built-in presets with natural-language trigger matching:

PresetTrigger wordsDefault team
Researchresearch, find, look up, compareLeader + 2 researchers
Developmentbuild, implement, code, add featureLeader + 2 developers
Analysisanalyze, review, inspect, auditLeader + 2 explorers
Contentwrite, document, README, blog postLeader + writer + reviewer
Product/PMPRD, strategy, roadmap, OKR, GTMLeader + PM + researcher

Presets are a starting point. The interview and environment scan shape the final team every time.

Shared memory

The team writes to .kkirikkiri/teams/{team_name}/ in your project root (session-scoped, no collision between concurrent sessions):

FilePurpose
TEAM_PLAN.mdTask plan, role assignments, goals
TEAM_PROGRESS.mdLive progress — completed and pending items
TEAM_FINDINGS.mdDiscoveries, dead ends (DEAD_ENDS)
report.mdCanonical final report for this session

Saved teams are stored cross-session under .kkirikkiri/shared/saved-teams/. If a team member is replaced mid-task, the new member reads these files and catches up immediately.

Validation loop

RoundStrategy
Round 1Original team executes
Round 2Auto-judge: keep (A) / full replacement (B) / partial swap (C)
Round 3Full team rebuild, unconditionally

Multi-model support

Claude + Codex CLI (code & large-scale analysis, cross-model review) + Antigravity CLI agy (design/UI) can each take different roles in the same team. kkirikkiri auto-detects what is installed and optimizes accordingly. Claude-only works fine if no external CLIs are present.

Agent auto-detection and reuse

If .claude/agents/ contains agent definitions, kkirikkiri detects them and recommends relevant ones per preset:

PresetExample agents
Researchdeep-research, data-analyst
Developmentcode-reviewer, architect
Analysiscode-analyzer, security-reviewer
Contentwriter, translator

After a successful run, you can save well-performing team members back to .claude/agents/ for reuse in other projects.

Spawn stability

If a team member fails to join:

  1. Retry once with the same configuration
  2. Retry with a downgraded model
  3. Continue with the remaining team members

Team save and reload

/kkirikkiri use the research team from last time

Requirements

Required

  • Claude Code (latest)
  • Agent Teams feature flag:
    // ~/.claude/settings.json
    {
      "env": {
        "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
      }
    }
    
  • Node.js (for external CLI integrations)
  • tmux (optional): only for split-pane team display. Teams run in-process without it. brew install tmux (macOS) / apt install tmux (Linux)

Optional (multi-model)

npm install -g @openai/codex                                    # Codex CLI — code & large-scale analysis, cross-model review
curl -fsSL https://antigravity.google/cli/install.sh | bash     # Antigravity CLI (agy) — design/UI (replaces Gemini CLI)

Works without these. Claude handles the full team on its own.

Cost reference

Team sizeEstimated timeCost level
2–3 members5–15 minLow
4–5 members10–30 minMedium
5+ members, multi-round30 min–1 hrHigh

Reduce team size or use Codex/Antigravity CLI to lower costs.


License

MIT


Like-minded agents, gathered for your goal.

Ver o original no GitHub

Bom para

Relacionados