Claudew — Agentes Claude Code em Sandbox Docker Paralelos
crionuke/claudew
Múltiplos agentes Claude Code paralelos em sandboxes Docker isolados
Ver no GitHubOrquestra múltiplos agentes Claude Code rodando em paralelo, cada um isolado em seu próprio sandbox Docker colorido com suas próprias ferramentas, regras e skills. Permite delegar tarefas cross-repo (editar múltiplos repositórios e abrir PRs) mantendo você no controle final de revisão e merge.
Instalação
npx convez-ai crionuke/claudew1 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
- Paralelizar tarefas do backlog entre agentes em clones independentes
- Edições cross-repo automáticas com abertura de PRs em cada repositório
- Executar migrações ou refatorações grandes em background enquanto você trabalha
- Scaffold de novos projetos seguindo convenções existentes
Veredito

Ferramenta especializada para times que precisam paralelizar tarefas de código com Claude, mantendo isolamento total entre agentes e controle humano garantido. Demanda familiaridade com Docker, shell e git; customização é esperada para seu stack específico.
README
A tradução para o português deste README ainda está na fila.
CLAUDEW
Built to collaborate, not to work autonomously.
Run multiple Claude Code agents in parallel, each sealed in its own color-coded
Docker sandbox loaded with your dev tools, rules, and skills. Because every
worker is a self-contained sandbox that can't touch your host or the others, you
can hand it --dangerously-skip-permissions and skip the permission prompts on
every step. You stay the one who reviews and merges — the sandbox just lets the
agent move without you gating each action. Each worker carries a multi-repo
workspace, so you can hand one agent a cross-cutting task and let it edit every
affected repo and open a PR in each. Launch and color-code them all from a
single open.sh.
This repo ships tuned to my own preferences and tech stack — Java/Quarkus rules, GitHub flow, iTerm2 launcher, the specific skills and coding standards in
./home/. None of it is load-bearing. Swap the home-skeleton files, drop or add skills, change the base image indocker/Dockerfileto whatever language and tooling you use. The core idea — multiple Claude Code agents running in parallel, each sealed in its own throwaway sandbox — and the advantages that come with it stay exactly the same whoever you are.
Use cases
- Parallelize one backlog — several agents on different issues in one repo, each with its own clone and PR.
- Hand off, then review — an agent goes all the way to a PR; you keep review and merge.
- Cross-repo changes — one agent edits every affected repo and opens a PR in each.
- Juggle several projects — one agent per repo, one tab each.
- Long background jobs — park a migration or big refactor in one agent while you build in another.
- Bootstrap new projects — an agent scaffolds a fresh repo following your existing conventions.
Usage
git clone https://github.com/crionuke/claudew.git
cd claudew
cp .env.sample .env # set GIT_USER_NAME, GIT_USER_EMAIL, GH_TOKEN, CONTEXT7_API_KEY
./open.sh -i # all agents, each in its own iTerm2 tab
open.sh builds and starts each worker on demand. -i opens every agent in its
own iTerm2 tab. To run a single agent in the current terminal, use -A, -B,
-C, or -D. Run with no argument to print usage.
Rules and skills
./home/ is a skeleton overlaid onto each worker's $HOME on start. It carries
the master ruleset, path-scoped rules, skills, and reference docs — everything
config-like lives there, while ~/workspace/ holds only your git clones. Drop
in your own files to change any of it; per-repo rules win on conflict.
Stack
- Ubuntu 24.04, Node.js 22, Claude Code
- Java 25 + 21 (Temurin, 25 default) + Quarkus CLI via SDKMAN
gh,dockerCLI (Compose + Buildx plugins)- Per-worker SSH key + git identity