convez.ai
MCP ServerAutomaçãoIntermediário
harun-yardimci

AgentFlow

harun-yardimci/agentflow

Orquestração de agentes IA com pipelines visuais e multi-LLM

Ver no GitHub
BomPluginScore · de 0 a 100

Plataforma de orquestração de agentes IA baseada em pipelines para executar tarefas multi-etapa com Claude, Codex, Gemini e Antigravity CLI. Oferece dashboard visual em tempo real, isolamento via git worktrees e controle programático via servidor MCP para coordenar fluxos complexos de trabalho.

Instalação

$npx convez-ai harun-yardimci/agentflow

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
9
Forks
3
Acompanhando
0
Saúde
100%
Último commit
há 1 mês
Versão
v1.4.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Automação de fluxos de trabalho que envolvem múltiplos modelos IA
  • Monitoramento em tempo real de pipelines de codificação e tarefas
  • Integração programática de agentes IA em aplicações via MCP server

Veredito

AgentFlow é um orquestrador sólido para quem precisa coordenar múltiplos agentes IA em pipelines estruturados. Bem mantido, com suporte a vários CLIs e MCP server integrado — ideal para cenários de automação complexa, embora exija familiaridade com Node.js e CLIs de IA instalados localmente.

README

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

AgentFlow

npm version License CI Node.js Version

Pipeline-based orchestration for AI coding agents. Build multi-step task graphs, run them with Claude Code, Codex, Gemini, or Antigravity CLI, and monitor the whole workflow from a visual dashboard or over MCP.

AgentFlow pipeline board — five stages running across Claude, Codex, and Gemini side-by-side, each task in its own git worktree

Run it

npx @argustech/agentflow

Open http://localhost:3100. No install, no config — npx pulls the package, builds the runtime, starts the dashboard on port 3100.

You'll also need at least one supported AI CLI on your machine (see Supported Agent CLIs below).

Highlights

  • Multi-agent execution for Claude Code, Codex CLI, Gemini CLI, and Antigravity CLI (agy)
  • Real-time dashboard for pipeline state, logs, outputs, and git diffs
  • MCP server for programmatic pipeline control from compatible clients
  • Git worktree isolation so tasks can run without stepping on each other
  • Interactive execution for approvals, follow-up questions, and tool access
  • Shared pipeline context with built-in persistence on SQLite
  • Optional Slack and Telegram notifications

Install

Requirements

  • Node.js ^20.19.0 or >=22.12.0
  • Git
  • At least one supported AI CLI installed locally

From npm (recommended)

# One-shot run, no install (slowest start, no global pollution)
npx @argustech/agentflow

# Global install — provides the `agentflow` command on PATH
npm install -g @argustech/agentflow
agentflow                  # start the dashboard
agentflow --help           # see all commands

A specific version pin:

npx @argustech/agentflow@1.0.0

From source

git clone https://github.com/harun-yardimci/agentflow.git
cd agentflow
npm install
npm run dev

This is the right path if you want to hack on AgentFlow itself, run the unbundled dev server with hot reload, or contribute back.

Storage

AgentFlow stores its shared SQLite database at ~/.agentflow/agentflow.db by default. Override with AGENTFLOW_DB_PATH. Worktrees, attachments, and uploads live under the same directory.

Supported Agent CLIs

Install at least one of these tools before running pipelines:

  • Claude Code: npm install -g @anthropic-ai/claude-code
  • Codex CLI: npm install -g @openai/codex
  • Gemini CLI: npm install -g @google/gemini-cli
  • Antigravity CLI (agy): Google's successor to the Gemini CLI — see antigravity.google. The Gemini CLI stops serving free/Google One tiers on June 18, 2026; agy is the migration path.

Antigravity note: agy emits plain-text output only (no token/cost metrics), and this build's --model flag accepts only Google's backend-defined model ids — unrecognized values fall back to the account default. AgentFlow seeds the Gemini 3 family best-effort; model switching firms up as Google stabilizes those ids.

Development Commands

Run from a source checkout (git clone + npm install). End users of the published package don't need any of these — they just npx @argustech/agentflow.

npm run dev          # frontend + backend with hot reload
npm run dev:fe       # frontend only
npm run dev:be       # backend only
npm run lint         # eslint + typecheck
npm run typecheck    # typecheck only
npm test             # vitest run
npm run build        # production bundle
npm run mcp          # MCP server over stdio
npm start            # production server (from source)

MCP Usage

AgentFlow exposes an MCP server so compatible assistants can create and manage pipelines programmatically.

Important runtime notes:

  • AGENTFLOW_PORT is the AgentFlow app port, not the internal dev API port
  • In development, the UI runs on AGENTFLOW_PORT and the backend on AGENTFLOW_PORT + 1
  • By default, npm run dev, npm start, and npm run mcp share the same DB unless AGENTFLOW_DB_PATH is overridden

Claude Code

# Globally installed
claude mcp add agentflow --env AGENTFLOW_PORT=3100 -- agentflow mcp

# Or via npx, no install needed
claude mcp add agentflow --env AGENTFLOW_PORT=3100 -- npx @argustech/agentflow mcp

Claude Desktop / Cursor

Add this to your MCP config file:

{
  "mcpServers": {
    "agentflow": {
      "command": "npx",
      "args": ["@argustech/agentflow", "mcp"],
      "env": {
        "AGENTFLOW_PORT": "3100"
      }
    }
  }
}

Available MCP Tools

ToolDescription
list_pipelinesList all pipelines with tasks and logs
get_pipelineGet pipeline details by ID
create_pipelineCreate a new pipeline with optional tasks
delete_pipelineDelete a pipeline
add_taskAdd a task to a pipeline
delete_taskDelete a task
approve_taskApprove a pending task
reject_taskReject a task
move_taskMove a task to a different status
complete_taskComplete a task and cascade to dependents
run_pipelineStart pipeline execution
list_agentsList available agents
update_agentUpdate agent configuration
get_logsGet pipeline activity logs
get_pipeline_contextRead shared pipeline context
set_pipeline_contextWrite to shared pipeline context
list_pending_questionsList pending interactive questions
respond_to_questionRespond to an interactive question

Architecture

+------------------+     +-----------------+     +------------------+
|   React Frontend | --> | Express Backend | <-- |   MCP Server     |
|   (Vite + TW v4) |     | (REST API)      |     | (stdio / SSE)    |
+------------------+     +--------+--------+     +--------+---------+
                                  |                        |
                           +------+------+                 |
                           |   SQLite    | <---------------+
                           +------+------+
                                  |
                    +-------------+-------------+
                    |             |             |
              +-----------+ +-----------+ +-----------+
              | Claude CLI| | Codex CLI | | Gemini CLI|
              +-----------+ +-----------+ +-----------+
  • Frontend: React 19, TypeScript strict mode, Tailwind CSS v4, Vite 7
  • Backend: Express 5, better-sqlite3, Zod validation
  • MCP: @modelcontextprotocol/sdk with stdio and SSE transports
  • Execution: worker pool that spawns CLI processes inside isolated git worktrees

Configuration

Environment variables are documented in .env.example.

VariableDefaultDescription
PORT3100AgentFlow app port
AGENTFLOW_PORT3100App port hint for MCP clients
AGENTFLOW_DB_PATH~/.agentflow/agentflow.dbOverride the shared SQLite DB path
TELEGRAM_BOT_TOKEN—Telegram notifications
TELEGRAM_CHAT_ID—Telegram destination chat
SLACK_WEBHOOK_URL—Slack webhook notifications
SLACK_BOT_TOKEN—Slack bot notifications
SLACK_CHANNEL—Slack channel target

Project Structure

src/             Frontend (React components, hooks, context)
server/          Backend (Express routes, services, execution engine)
  db/            SQLite schema, connection, seed
  engine/        Task runner, worker pool, worktree manager
  executor/      CLI executor, templates, output parser
  routes/        REST API endpoints
  services/      Business logic
  safety/        Output safety checks
mcp/             MCP server
bin/             CLI entry point
tests/           Unit and integration tests

Community

License

AgentFlow core is licensed under Apache-2.0.

Ver o original no GitHub

Bom para

Relacionados