Kobe
Sma1lboy/kobe
Terminal IDE para rodar agentes de IA em paralelo com worktrees isoladas
Ver no GitHubIDE de terminal para executar múltiplos agentes de IA em paralelo, cada um isolado em seu próprio git worktree. Reimagina o fluxo de trabalho com agentes IA (Claude, Codex, Copilot) através de uma interface TUI persistente, SSH-friendly, ideal para devboxes e ambientes remotos.
Instalação
npx convez-ai Sma1lboy/kobe1 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
- Executar múltiplas tentativas de codificação IA simultaneamente
- Desenvolvimento remoto em devboxes e VPS via SSH
- Orquestração de agentes com fan-out scriptável
- Comparação de abordagens diferentes para um mesmo problema
Veredito

Ferramenta robusta e inovadora que resolve um problema real: gerenciar múltiplos agentes IA de forma segura e paralela. Excelente para quem trabalha em ambientes remotos e quer experimentação séria, não casual. Requer setup técnico (Bun, tmux, CLI de IA) mas oferece valor genuíno em produtividade.
README
A tradução para o português deste README ainda está na fila.
A quick look
https://github.com/user-attachments/assets/b37c5199-9a16-4a31-8ac6-79e968658867
kobe opens into a tmux workspace with:
- Tasks - create, switch, archive, rename, retarget.
- Engine - the live AI CLI session.
- Ops - changed files, previews,
@filementions, PR prompts. - Shell - a normal shell inside the task worktree.
AI agents are useful one at a time. kobe is for when you want five attempts running at once.
Task = git worktree + tmux session + branch
Create a task, send it to claude, codex, or copilot, detach, reattach, compare the worktree, keep the good branch, archive the rest. It runs where your code already lives: your laptop, a devbox, a VPS, or any machine you can SSH into.
ssh devbox
cd repo
kobe
Why try it
- Made for SSH/devboxes - no browser, VNC, or desktop app; the terminal is the product.
- Persistent by default - agents live in tmux, so disconnects do not kill the work.
- Safe parallelism - every attempt gets its own branch and worktree.
- Real environment - agents run next to your dependencies, services, credentials, and build cache.
- Scriptable fan-out -
kobe apilets another agent or shell script spawn more tasks.
Install
Requirements: Bun >= 1.3.11, tmux, and at least one engine CLI on PATH (claude, codex, or copilot).
bun install -g @sma1lboy/kobe
kobe
Or:
bunx @sma1lboy/kobe
First task: press n, choose a repo/base branch/engine, then prompt the engine pane. kobe creates the worktree under:
<repo>/.claude/worktrees/<task-slug>/
Useful keys
| Key | Action |
|---|---|
ctrl+h/j/k/l | Move between Tasks, engine, Ops, and shell panes. |
ctrl+q | Detach; tasks keep running in tmux. |
ctrl+t | New ChatTab on the same task/worktree. |
ctrl+[ / ctrl+] | Previous / next ChatTab. |
F2 | Rename the current ChatTab. |
tmux prefix f | Open the new-task dialog. |
More: docs/KEYBINDINGS.md.
Fan out
kobe api fan-out \
--repo "$PWD" \
--agents claude:2,codex:1 \
--prompt "Try three approaches to simplify the auth flow."
Install the companion skill so Claude Code knows when to use kobe api:
npx skills add Sma1lboy/kobe --skill kobe --agent claude-code
If it gets stuck
kobe doctor # read-only diagnosis
kobe reset # reset daemon + kobe tmux sessions; does not delete worktrees
Develop
bun install
bun run dev:sandbox
bun run typecheck
bun run lint
bun run test
Start with HANDOFF.md, docs/ARCHITECTURE.md, and packages/kobe/CHANGELOG.md.