convez.ai
AgenteAutomaçãoIntermediário
superliaye

My Agent Kits

superliaye/my-agent-kits

Wizard para setup rápido de agentes Claude e Codex com presets configuráveis

Ver no GitHub
InicialPluginScore · de 0 a 100

Kit de configuração para bootstrapping e atualização de agentes AI pessoais para Claude Code e Codex CLI. Automatiza a instalação de presets de skills e artefatos sem dependências de APM, oferecendo fluxo interativo ou one-shot com flags.

Instalação

$npx convez-ai superliaye/my-agent-kits

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
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 1 semana
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Inicializar ambiente de desenvolvimento com agentes pré-configurados
  • Alternar entre presets (engineering, productivity, feature-loop) sem reconfigurações manuais
  • Manter sincronizados os agentes globais com atualização única de comando

Veredito

Ferramenta prática para developers que usam Claude Code ou Codex, eliminando fricção no setup inicial de agentes. Arquitetura limpa (Node 20+, CLI puro) e bem documentada, mas escopo bastante específico ao ecossistema local da Codex.

README

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

my-agent-kits

Personal AI agent artifact wizard for Claude Code and Codex CLI. Bootstrap or update any repo's agent setup with one command.

Install (one time per machine)

git clone https://github.com/superliaye/my-agent-kits.git ~/my-agent-kits
cd ~/my-agent-kits
bash bootstrap.sh

Requires Node 20+ and Git. No APM CLI dependency. bootstrap.sh only installs Node deps — it does not touch your PATH or shell rc files.

Usage

Invoke the launcher from the kit directory (or anywhere via its absolute path) — every run deploys to your global agent directories, never into a repo:

cd ~/my-agent-kits
./bin/agent-kit init --default   # install recommended defaults, no prompts
./bin/agent-kit init             # interactive 5-step wizard
./bin/agent-kit update           # re-deploy the kit to your global install
./bin/agent-kit help

You can also invoke it by absolute path from anywhere:

~/my-agent-kits/bin/agent-kit init

To skip every prompt and install the recommended defaults (the pre-checked presets and the preset's agents), pass --default — the "enter through everything" path. Explicit flags still override individual defaults:

./bin/agent-kit init --default                       # zero prompts, all defaults
./bin/agent-kit init --default --preset productivity # defaults, but a different preset

One-shot install (each flag is one decision)

Passing --preset and --agents together is enough — the wizard treats it as "you've decided" and skips all prompts:

./bin/agent-kit init --preset engineering --agents claude

That concatenates the instructions inline into ~/.claude/CLAUDE.md and (with --agents codex) ~/.codex/AGENTS.md, and copies skills to ~/.claude/skills/ and ~/.agents/skills/. No apm.yml, no apm_modules/.

Common variations — change exactly the flag(s) that differ:

./bin/agent-kit init --preset engineering --agents claude,codex  # Codex too
./bin/agent-kit init --preset productivity --agents claude        # productivity preset

Flag reference:

  • --default — accept every wizard default (pre-checked presets and the preset's agents) and apply without prompting; explicit flags still override individual defaults
  • --preset NAME[,NAME2] — one or more of {engineering, experimenting-engineering, productivity, experimenting-productivity, feature-loop, loop-full-swe, financial, none}. Comma-separated names merge primitives (union, deduped per type); interactive form uses a multiselect prompt
  • --agents claude[,codex] — which agents to deploy to
  • --primitives '+name,-name' — tweak the preset's primitive set on the fly
  • --bundles name1,name2 — external installers (e.g. gstack) to run after primitives deploy. Always install globally. Pass --bundles '' to skip all.

Updating:

agent-kit update is a stateless global re-deploy. There is no per-repo state to diff against — it re-resolves the working set the same way init --default does (same --preset/--agents shape) and re-runs the idempotent deploy, so the global install matches the current kit:

./bin/agent-kit update                                  # re-resolve defaults, re-deploy globally
./bin/agent-kit update --preset engineering --agents claude  # re-deploy a specific preset/agent set

What's in here

PathPurpose
presets/*.yamlBundled artifact selections (engineering, experimenting-engineering, productivity, experimenting-productivity, feature-loop, loop-full-swe, financial, none). Multi-select via --preset a,b
.apm/instructions/*.instructions.mdAlways-loaded rules, concatenated at deploy into ~/.claude/CLAUDE.md (inline) and ~/.codex/AGENTS.md
.apm/skills/<name>/SKILL.mdReusable workflows — slash commands and multi-step skills. Authored in Claude format with disable-model-invocation: true for manual-only.
.apm/plugins/*.plugin.mdClaude Code plugin pointers (e.g., superpowers)
.apm/bundles/*.bundle.mdExternal installers wrapped as deployable primitives (e.g., gstack — 30+ /gstack-* skills; hyperframes — HTML video rendering). Two installer.kind flavors: setup-script (clone + run) and npx-skills (npx skills add <pkg>). Always installed globally; common runtime deps auto-installed by the wizard. See docs/maintaining-bundles.md.
bin/agent-kitLauncher — invoke as ./bin/agent-kit from the kit dir, or via absolute path from anywhere
lib/wizard.js + lib/*.jsWizard implementation (Node 20+)
test/Docker-based test matrix

What lands where

agent-kit init writes the artifacts to your global agent directories — nothing lands in any repo. There is no repo-local state file:

~/.claude/CLAUDE.md     # instructions concatenated inline (overwritten each run)
~/.claude/skills/       # Claude Code reads skills here
~/.codex/AGENTS.md      # instructions concatenated inline (if --agents codex)
~/.agents/skills/       # cross-client skills (Codex reads here; if --agents codex).
                        #   Each Codex skill also gets a manual-only
                        #   ~/.agents/skills/<name>/agents/openai.yaml sidecar.

That's it. No apm.yml, no apm_modules/, no per-rule .claude/rules/*.md files, no per-repo state file, and no instructions or skills copied into the repo itself.

Tests

Default — isolated Docker run (Unix-only validation, doesn't touch your host's ~/.claude/ or ~/.codex/):

npm test
# equivalent to: docker build -q -f test/Dockerfile.test -t my-agent-kits-test . && docker run --rm my-agent-kits-test

Opt-in — run the suite on this machine (faster inner loop, but cases overwrite your real ~/.claude/CLAUDE.md / ~/.codex/AGENTS.md and may delete ~/.claude/plugins/):

npm run test:host
# or via the CLI:
./bin/agent-kit test --host

Host mode is gated behind AGENT_KIT_TEST_HOST=1 so direct bash test/run-tests.sh runs refuse without the env var; use the npm script or the --host flag.

Design history (dated archives)

These are point-in-time records of the original installer design, kept for context. They describe the architecture as it stood when written (May 2026) and predate the global-only refactor — they are not a description of current behavior. For current behavior, see the sections above.

Ver o original no GitHub

Bom para

Relacionados