convez.ai
AgenteAutomaçãoIntermediário
agent-tower

Agent Tower

agent-tower/core

Dashboard unificado para coordenar múltiplos agentes de IA com isolamento Git automático

Ver no GitHub
BomPluginScore · de 0 a 100

Central de comando unificada para gerenciar múltiplos agentes de IA (Claude Code, Gemini CLI, Cursor Agent, Codex) em paralelo. Fornece um painel único para controlar tarefas, isola automaticamente branches Git, permite escolher providers por tarefa para economizar tokens e oferece acesso remoto via celular.

Instalação

$npx convez-ai agent-tower/core

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
402
Forks
21
Acompanhando
17
Saúde
37%
Último commit
há 5 dias
Versão
v0.5.3-beta.12
Issues abertas
0
Contribuidores
4

Casos de uso

  • Executar múltiplas tarefas de IA em paralelo sem conflitos de código
  • Gerenciar custos escolhendo modelos diferentes por tarefa (Opus para planejamento, Codex para execução, etc)
  • Monitorar progresso de agentes remotamente via celular com notificações em tempo real
  • Revisar e mesclar alterações de código com diff viewer integrado

Veredito

Solução prática que resolve problemas reais de quem trabalha com múltiplos agentes de IA simultaneamente. Arquitetura bem pensada com Git Worktree automático e suporte a MCP, mas exige Node.js e familiaridade com terminal. Recomendado para desenvolvedores que querem maximizar produtividade e controlar custos.

README

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

English | 简体中文

Agent Tower

Too many terminal sessions (Claude Code / Codex ...) open? Here's their command center.

Why I Built This

When I first started using Claude Code, I'd open one terminal and sit there, bored, watching it spit out characters. Then I got clever — opened multiple terminals, ran different tasks at the same time, even worked on different projects simultaneously. Productivity went through the roof. I felt like an absolute genius.

That lasted about two days before the problems hit:

  • Visual chaos: My desktop was covered in Claude Code terminals. I constantly lost track of which window was running which task.
  • Edit conflicts: Multiple tasks touching the same files meant merge conflicts everywhere. I tried Git Worktree to isolate each task, but manually running commands to split, rebase, and merge was still a pain.
  • No mobile access: Sitting in front of the computer staring at terminals gets old. When I'm chilling on the balcony with my phone, why can't I check on how my AI workhorses are doing?
  • Token bills: More tasks meant bigger bills, and it stung. Plenty of simple tasks could run on cheaper models, but manually switching configs was a hassle — even with tools like ccswitch, you have to wait for the current task to finish before switching.

So Agent Tower was born out of necessity — a single dashboard that brings all your agent tasks, terminals, and code changes into one interface. Auto-creates isolated branches, lets you pick providers per task, works from your phone, and notifies you when it's time to review.

Core Features

🎯 One Dashboard for All Agents

No more juggling terminal windows. All projects, all tasks, all agents — one page. Create a task, pick an agent, hit start — output, progress, and code changes are visible in real time. When a task finishes, it automatically moves to "Ready for Review". You decide whether to merge.

🔀 Automatic Git Worktree Isolation

Each task gets its own Git branch automatically. Agents work in isolated environments, eliminating code conflicts at the root. One-click merge back to main, with line-by-line review in the diff viewer.

💰 Pick Providers Per Task, Save Money Effortlessly

Each task can use a different provider. Throw translation tasks to MiniMax, planning to Opus, execution to Codex — each in their lane, no waiting for one task to finish before switching. Agent Tower helps you spend smart.

📱 Monitor Progress from Your Phone

One-click Cloudflare tunnel for remote access — open the dashboard right in your phone's browser. Away from your desk? Still in the loop. Task done? Desktop notification or Lark webhook lets you know it's review time.

🤖 Supports Major AI Agents

  • Claude Code · Gemini CLI · Cursor Agent · Codex

No vendor lock-in. Use whichever agent you prefer. Each supports custom Profile variants.

📡 MCP Protocol Integration

Built-in MCP server lets agents read the task board, claim tasks, and report progress directly. You're not just managing agents — they can collaborate proactively.

Getting Started

Recommended: global install (simplest — one command and you're done)

npm install -g agent-tower
agent-tower

Open http://localhost:12580 and start using it.

Prerequisite: Node.js >= 18

Configure MCP (Optional)

Let Claude Code operate the task board directly:

{
  "mcpServers": {
    "agent-tower": {
      "command": "agent-tower-mcp",
      "args": []
    }
  }
}

Development from Source

git clone https://github.com/agent-tower/core.git
cd agent-tower
pnpm setup          # Install deps + build shared package
pnpm dev            # Start all services in dev mode

Architecture Overview

┌─────────────────────────────────────────────────┐
│                  Browser (React)                 │
│  Kanban ─ Terminal ─ Code Editor ─ Git Changes   │
│  TanStack Query (server cache) + Zustand (UI)   │
└──────────────────┬──────────────────────────────┘
                   │ HTTP REST + Socket.IO (/events)
┌──────────────────┴──────────────────────────────┐
│                Fastify Server                    │
│  ┌───────────┐ ┌──────────┐ ┌────────────────┐  │
│  │ REST API  │ │ Socket.IO│ │  MCP Server    │  │
│  │ (16 routes)│ │ (realtime)│ │ (agent integ.) │  │
│  └─────┬─────┘ └────┬─────┘ └───────┬────────┘  │
│        └──────┬──────┘               │           │
│        ┌──────┴──────┐               │           │
│        │  Services   │               │           │
│        │ Session Mgmt│               │           │
│        │ Workspace   │               │           │
│        │ Git/Worktree│               │           │
│        │ Notif/Tunnel│               │           │
│        └──────┬──────┘               │           │
│        ┌──────┴──────┐               │           │
│        │AgentPipeline│               │           │
│        │ PTY + Parser│               │           │
│        │ + MsgStore  │               │           │
│        └──────┬──────┘               │           │
│               │ node-pty             │           │
│        ┌──────┴──────┐               │           │
│        │  Executors  │               │           │
│        │ Claude Code │               │           │
│        │ Gemini CLI  │               │           │
│        │ Cursor Agent│               │           │
│        │ Codex       │               │           │
│        └─────────────┘               │           │
│                                      │           │
│        ┌─────────────┐               │           │
│        │ SQLite      │◄──────────────┘           │
│        │ (Prisma ORM)│                           │
│        └─────────────┘                           │
└─────────────────────────────────────────────────┘

Design Philosophy

Agents as Team Members: Agent Tower treats AI agents like team members. Projects are workspaces, tasks are work items, agents are executors. Manage assignments via the kanban board, monitor execution through terminals, review results through Git integration.

Pipeline Architecture: Each agent session is managed by an AgentPipeline with three core components:

  • PTY: Pseudo-terminal for interacting with the agent CLI process
  • Parser: Output parser that structures raw agent output into tool calls, code changes, etc.
  • MsgStore: Message store using JSON Patch for incremental sync to the frontend

Worktree Isolation: Each task creates an independent Git Worktree, so agents work on isolated branches — preventing conflicts when multiple agents modify code simultaneously.

Tech Stack

LayerTechnology
FrontendReact 19 + Vite 7 + TypeScript 5
StylingTailwindCSS v4 + shadcn/ui (Radix UI)
StateTanStack Query v5 + Zustand v5
Terminalxterm.js 5 + Monaco Editor
BackendFastify 4 + Socket.IO 4
DatabasePrisma 5 + SQLite
Processnode-pty
ProtocolMCP (Model Context Protocol)
Packagepnpm monorepo

Contact

If you have any questions or suggestions, feel free to add me on WeChat:

License

Licensed under the Apache License, Version 2.0. See LICENSE for details.

Ver o original no GitHub

Bom para

Relacionados