convez.ai
MCP ServerDevOps & InfraestruturaAvançado
ashuiGordon

Argo - Plataforma de Colaboração Multi-Agent Local-First

ashuiGordon/argo

Orquestre equipes de agentes de código com segurança, visualização e deploy automático.

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP que orquestra múltiplos agentes de IA (Claude Code, Codex) como uma equipe colaborativa com interface visual, aprovações de segurança e deploy one-click. Executa tudo localmente sem dependências de nuvem, oferecendo isolamento via git worktree, classificação de risco em 4 níveis e 8 agentes especializados pré-configurados (Arquiteto, Planejador, Executor, Revisor, Debugger, Testador, Designer, Ops).

Instalação

$npx convez-ai ashuiGordon/argo

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
3
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 2 meses
Versão
v0.1.0
Issues abertas
0
Contribuidores
3

Casos de uso

  • Desenvolvimento colaborativo onde múltiplos agentes trabalham em paralelo em diferentes aspectos (arquitetura, testes, reviews)
  • Pipelines de código com aprovações de segurança automáticas para operações críticas
  • Deploy one-click de aplicações web geradas por IA para Vercel, Netlify, Docker ou Fly.io

Veredito

Ferramenta robusta e arquitetura ambiciosa para automação de fluxos de desenvolvimento com múltiplos agentes. Requer compreensão profunda de orquestração de IA e operações de infraestrutura, mas oferece controle total local e segurança granular — ideal para equipes que querem governança forte sobre agentes de código.

README

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

Argo

English | 中文

A local-first multi-agent programming collaboration platform. Argo wraps AI coding CLI tools (Claude Code, Codex) as visual agents and provides a unified interface to orchestrate them working together on your local projects.

Argo Overview

What is Argo?

Unlike traditional AI coding assistants that call LLM APIs directly, Argo runs Claude Code CLI / Codex CLI as subprocess agents via SDK async generators. Each agent inherits the full CLI capabilities (file operations, command execution, code search) while Argo adds visualization, team orchestration, and safety controls on top.

Core capabilities:

  • Visual Conversations — Streaming display of agent thinking, tool calls, and artifacts, replacing the raw terminal experience
  • Multi-Agent Teams — 8 specialized built-in agents + custom agents, auto-orchestrated by a Moderator with parallel execution + git worktree branch isolation
  • Safety Approvals — 4-level risk classification (Low/Medium/High/Critical), real-time approval cards for dangerous operations, 30s auto-deny timeout
  • Context Memory — Auto session summaries + message pinning + context injection for cross-session continuity
  • One-Click Deploy — Code generated in conversations can be previewed locally, deployed to Vercel/Netlify/Docker/Fly.io, or packaged for download
  • MCP Extensions — Extend agent capabilities via Model Context Protocol servers (GitHub, databases, search, browser, etc.)

Single Agent Chat

Interact with any specialized agent in a focused conversation. Streaming responses with tool-use visualization and approval cards.

Single Agent Chat

Architecture

Frontend (localhost:5173)
    │  React 18 + Vite + Zustand + TailwindCSS
    │
    ▼  WebSocket + REST
Daemon (localhost:54321)
    │  Hono + Node.js + SQLite (WAL)
    │
    ▼  Claude Code SDK / Codex SDK
Agent CLI Subprocesses

All data stored locally in SQLite. No cloud dependency.

Built-in Agents

AgentRoleModelPermissions
ArchitectCode analysis, bug diagnosis, architecture suggestionsOpusRead-only
PlannerRequirements analysis, task decomposition, work planningOpusRead-only
ExecutorCode implementation with minimal diffsSonnetFull
ReviewerSpec compliance, security, code quality reviewOpusRead-only
DebuggerSystematic bug investigation and fixesSonnetFull
TesterComprehensive test writing (happy path + edge cases)SonnetFull
DesignerUI/UX implementation, responsive design, accessibilitySonnetFull
OpsGit operations, branch management, CI/CD, deploymentsSonnetFull

Agents are fully customizable — create your own with custom system prompts, model selection, tool permissions, MCP servers, and skills.

Agent Configuration

Team Collaboration

When multiple agents are assigned to a conversation:

  1. User sends a message
  2. Moderator analyzes the task and @mentions the needed agents
  3. Agents run in parallel (up to 5 concurrent), each in isolated git worktrees
  4. Results are merged and synthesized into a unified response
  5. Up to 3 rounds of agent collaboration per user message

Team Presets:

  • Auto — Moderator has full autonomy to decide which agents to invoke
  • Feature — Structured pipeline: Clarify → Plan → Implement → Test → Review → Commit

Team Collaboration

Prerequisites

  • Node.js >= 22.0.0
  • pnpm >= 9
  • Claude Code CLI (authenticated via OAuth)
  • Git (required for multi-agent parallel worktrees)
  • Codex CLI (optional alternative runtime)

Getting Started

# Clone the repository
git clone https://github.com/ashuiGordon/argo.git
cd argo

# Install dependencies
pnpm install

# Start development servers (daemon + UI)
pnpm dev

Once running:

On first visit, a development user is auto-created — no registration needed.

Project Structure

packages/
  shared/     # Shared types, schemas, constants
  daemon/     # Backend: API routes, orchestrator, session management,
              # approval system, event bus, memory, deploy manager
  ui/         # Frontend: React SPA with chat, workspace panel,
              # agent management, approval cards, deploy UI
avatars/      # Agent avatar images

Scripts

pnpm dev        # Start daemon + UI in development mode
pnpm build      # Build all packages for production
pnpm test       # Run tests
pnpm typecheck  # TypeScript type checking
pnpm lint       # ESLint
pnpm format     # Prettier formatting

Environment Variables

VariablePurposeRequired
ANTHROPIC_API_KEYSession summary generation (Claude API)Optional*

* Agent execution itself doesn't need this — Claude Code CLI manages its own authentication. Without the key, summaries fall back to a local extraction method.

Key Features

MCP Server Integration

Agents can be extended with MCP (Model Context Protocol) servers for additional capabilities:

  • 13 pre-configured servers (filesystem, git, github, postgres, sqlite, brave-search, puppeteer, slack, memory, fetch, etc.)
  • Install from MCP Marketplace or search the MCP Registry
  • Up to 20 MCP servers per agent

Skills System

Reusable prompt templates injected into agent system prompts:

  • 10 built-in general skills (code-review, write-tests, refactor, debug, etc.)
  • 4 deployment skills (preview, static, container, package)
  • Up to 50 skills per agent, max 10,000 characters each

Approval System

All agent tool calls flow through a risk classifier:

  • Auto-allow (Low): Read, Glob, Grep, WebSearch, etc.
  • Medium: File writes/edits, general Bash commands
  • High: Operations on sensitive files (.env, credentials)
  • Critical: rm -rf, DROP TABLE, git push --force, etc.

"Always Allow" rules can be configured per user to streamline repeated safe operations.

Deployment

Four deployment types available from within conversations:

  • Preview — Local dev server with framework auto-detection
  • Static — Deploy to Vercel or Netlify
  • Container — Docker build + Fly.io deployment
  • Package — Download as tar/zip archive

License

MIT

Ver o original no GitHub

Bom para

Relacionados