convez.ai
AgenteAutomaçãoAvançado
ujm

Conductor - Orquestração de Agentes IA

ujm/Conductor

Cockpit visual para orquestrar agentes IA com controle humano em cada etapa

Ver no GitHub
InicialPluginScore · de 0 a 100

Plataforma de orquestração que permite aos engenheiros dirigir e supervisionar múltiplos agentes de IA (como Claude Code) através de uma interface visual. Combine agentes em pipelines, gerencie dependências, defina fluxos de aprovação humana e monitore logs e mudanças de arquivos em tempo real.

Instalação

$npx convez-ai ujm/Conductor

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 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Automação de tarefas de desenvolvimento com revisão humana
  • Orquestração de pesquisa e análise envolvendo múltiplos agentes
  • Pipelines de processamento de dados com aprovações intermediárias
  • Gerenciamento de projetos técnicos complexos com IA assistida

Veredito

Conductor é uma solução sólida para times que querem aproveitar agentes IA em tarefas reais sem abrir mão do controle humano. Arquitetura clara (React + Express + WebSocket), features maduras (pipelines, aprovações, monitoramento), mas exige familiaridade com CLI/APIs e configuração YAML para máximo valor.

README

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

Conductor

日本語

An orchestration UI for humans to direct AI agents (Claude Code, etc.).


Overview

Conductor is a tool that combines multiple AI agents into pipelines, letting a human manage and supervise their progress. From the UI you define when, in what order, and under what conditions agents run, and you can monitor their output, logs, and file changes in real time while they execute.

This is aimed at engineers who want to use AI agents for real development, research, and automation tasks, but want a human in the loop because fully autonomous execution feels too risky.


Features

Pipeline Management (Pipeline View)

  • Drag and drop agent nodes to reorder and define execution order
  • Configure dependencies between nodes (on:done / on:error / on:approve / parallel)
  • Run or stop the entire pipeline at once, or execute individual nodes
  • Real-time display of each agent's state (idle / running / done / error / awaiting approval)
  • Progress bar and partial output shown on cards while running

Task Management (Task Board)

  • Manage tasks on a Kanban board (Todo / In Progress / Review / Done / Blocked)
  • Drag and drop task cards to change their status
  • Color-coded task priority (low / medium / high / critical)
  • Support for assigning tasks to agents

File Management (File Manager)

  • Browse the file tree under the project root
  • View, edit, and save text files in place
  • Edit .conductor/ configuration files (YAML) directly in the browser

Log Viewer

  • Real-time streaming of agent stdout/stderr
  • Filter logs by agent
  • Color-coded log levels (info / warn / error / debug)
  • Persisted to disk in JSONL format (organized by date)

Approval Queue

  • Nodes with an on:approve dependency block execution until a human approves
  • Approve or reject with a single button click
  • Notification badge shown in the header and navigation when approvals are pending

Agent Configuration

  • Register CLI agents (command, arguments, working directory) from the UI
  • Also supports registering REST API agents (Base URL)
  • Configure timeout and retry count individually per agent

Architecture

Frontend (React + Vite, :5173)
  └── useWebSocket hook (single WebSocket connection)
        ├── Incoming events → Zustand stores (auto-update)
        └── Outgoing messages → Backend

Backend (Express 5 + ws, :3001)
  ├── REST API  /api/*  ← Agent / pipeline / task / file management
  └── WebSocket /ws     ← Broadcasts state changes, output, and logs

Key Components

LayerClass / ModuleRole
ServiceOrchestratorServicePipeline state machine. Dependency resolution and retry control
ServiceApprovalGatewayBlocks on:approve nodes via Promises
ServiceLogServiceJSONL file writes + WebSocket streaming
ServiceFileWatcherWatches file changes with chokidar
ConnectorCliAdapterLaunches CLI agents via child_process.spawn
ConnectorRestAdapterCalls REST agents via fetch + AbortController
FrontendZustand storesThe single source of state, updated directly by WebSocket events

Configuration Files

All configuration is persisted as YAML under the .conductor/ directory.

.conductor/
├── project.yaml          # Project name and root path
├── pipeline.yaml         # Agent order and dependencies
├── agents/
│   └── claude-code.yaml  # Agent connection settings (one file per agent)
└── tasks/
    └── <uuid>.yaml       # Tasks (one file per task)

Example pipeline.yaml:

version: 1
name: My Pipeline
agents:
  - id: node-01
    agent: claude-code
    order: 1
    task: "Update the README"
    instruction_files: []
  - id: node-02
    agent: claude-code
    order: 2
    task: "Run the tests"
    depends_on:
      - agent: node-01
        trigger: done       # Starts automatically once node-01 completes

Trigger types:

ValueMeaning
doneStarts automatically when the previous node completes successfully
errorStarts automatically when the previous node ends with an error
approveWaits until a human clicks the approve button
parallelNo dependency (starts at the same time as the pipeline)

Example Agent Configuration File (agents/*.yaml)

id: claude-code
name: Claude Code
type: cli
icon: "🤖"
color: "#4f8ef7"
connection:
  command: claude
  args:
    - "--dangerously-skip-permissions"
  cwd: "{project_root}"
defaults:
  timeout_minutes: 30
  retry_count: 2
  approval_required: false
  context_files:
    - memory/project-context.md   # File appended to the prompt before execution

{project_root} and {env.VAR_NAME} are resolved by the backend at startup.


Usage

Prerequisites

  • Node.js 20 or later
  • Claude Code CLI installed (the claude command must be available)

Installation

git clone https://github.com/ujm/Conductor.git
cd Conductor

# Install backend dependencies
cd backend && npm install && cd ..

# Install frontend dependencies
cd frontend && npm install && cd ..

Running

# Set PROJECT_ROOT to the path of the project you want to manage
PROJECT_ROOT=/path/to/your/project npm run dev

Open http://localhost:5173 in your browser.

Note
npm run dev starts both the backend (:3001) and frontend (:5173) at the same time.
Frontend requests to /api and /ws are forwarded to the backend via the Vite dev proxy.

Initial Setup

  1. The .conductor/ directory is created automatically if it doesn't exist
  2. Register agents from the Agent Config screen, or create YAML files directly under .conductor/agents/
  3. Add nodes and configure dependencies in Pipeline View
  4. Run the pipeline using the Run button in Pipeline View

Running a Pipeline

  1. Open Pipeline View
  2. Press the ▶ Run button on an agent card to run it individually, or Run Pipeline in the header to run the whole pipeline
  3. Watch real-time output in the Log Viewer
  4. Approve or reject on:approve nodes from the Approval Queue screen

Assigning Tasks to Agents

  1. Create a new task on the Task Board (Enter key or + button)
  2. Set the assigned_agent field in .conductor/tasks/<uuid>.yaml
  3. Reference that task YAML in instruction_files in the pipeline node configuration

Testing

cd backend && npm test

Tech Stack

AreaTechnology
BackendNode.js 20 / TypeScript 6 / Express 5 / ws
FrontendReact 19 / Vite 8 / TypeScript 6 / Tailwind CSS v4
State ManagementZustand 5
Pipeline UIReact Flow (@xyflow/react)
TestingVitest
Config FormatYAML (js-yaml)
Ver o original no GitHub

Bom para

Relacionados