nodejs-swarm-openai

nodejs-swarm-openai is an experimental Node.js SDK for orchestrating multi-agent conversations using OpenAI's models (GPT-4o, GPT-4, GPT-3.5). This library is a custom test implementation built from scratch, featuring support for finalizer agents, async tools, and improved error handling. It is intended for learning, prototyping, and exploring the principles behind Swarm-like coordination.

✅ NPM

• 📦 View on npm

✅ Features

• ✨ Orchestrates multiple agents (chatbots) with distinct roles and instructions.
• 🔁 Automatic tool call handling via OpenAI function calling.
• 🧠 Optional finalizer agent to generate a unified final message.
• 🔧 Supports async tool functions (e.g., DB queries).
• 🐛 Safe internal error handling (errors never leak to user).
• 🧪 Built-in debug mode for developer insight.
• 🌐 Proxy support.

🔧 Installation

npm install nodejs-swarm-openai

📦 Usage

import { Agent, Swarm, printMessagesPretty, getFinalAnswer } from 'nodejs-swarm-openai';

const needsAnalyst = new Agent({
  name: 'needs_analyst',
  description: 'Understands customer needs',
  instructions: 'Identify what matters most to the customer: budget, quality, speed, etc.',
  model: 'gpt-4o',
  tools: []
});

const positioningAgent = new Agent({
  name: 'positioning',
  description: 'Describes company strengths',
  instructions: 'Mention experience, materials, guarantees, and successful projects.',
  model: 'gpt-4o',
  tools: []
});

const sellerAgent = new Agent({
  name: 'seller',
  description: 'Invites to action',
  instructions: 'Encourage user to leave a phone number or schedule a call.',
  model: 'gpt-4o',
  tools: []
});

const routerAgent = new Agent({
  name: 'manager',
  description: 'Manages tasks and delegates to agents',
  instructions: `
You are a Swarm manager. Given a customer request, decide which agent to invoke:

- needs_analyst — to analyze customer needs;
- positioning — to describe the company;
- seller — to invite to next step.

Always respond with tool_call only. Never reply directly.`,
  model: 'gpt-4o'
});

const swarm = new Swarm(process.env.OPENAI_API_KEY, {
  proxyUrl: process.env.PROXY_GPT // Optional proxy support
});

(async () => {
  const { messages } = await swarm.run({
    routerAgent,
    agents: [needsAnalyst, positioningAgent, sellerAgent],
    finalizerAgent: new Agent({ //Optional
    name: 'my_finalizer',
    description: 'Custom summary',
    instructions: 'Merge replies into 1 final answer. Use polite tone.',
    model: 'gpt-4o',
    tools: []
  }),
    messages: [
      { role: 'user', content: 'I want a cheap and high-quality wooden sauna' }
    ],
    maxTurns: 5,
    debug: true // Optional
  });

  printMessagesPretty(messages);
  const final = getFinalAnswer(messages);
  console.log('\nFinal Answer:\n', final || 'No final message.');
})();

✅ Use Cases

• Multi-step sales dialogs
• Lead qualification
• Information extraction and synthesis
• Decision trees based on roles

🛠 Advanced

Finalizer agent

By default, the library adds a final agent to summarize all assistant replies into one. You can override it:

const swarm = new Swarm(apiKey);
swarm.run({
  finalizerAgent: new Agent({
    name: 'my_finalizer',
    description: 'Custom summary',
    instructions: 'Merge replies into 1 final answer. Use polite tone.',
    model: 'gpt-4o',
    tools: []
  }),
  ...
});

Spent tokens

const result = await swarm.run({
  routerAgent,
  agents: [needsAnalyst, positioningAgent, sellerAgent],
  finalizerAgent,
  messages: [
    { role: 'user', content: 'I want a cheap and high-quality wooden sauna' }
  ],
  maxTurns: 6,
  debug: false
});
console.log(result.totalTokens)

{
  messages:[...]
  tokenUsage: {
    manager: 1797,
    needs_analyst: 262,
    positioning: 415,
    seller: 579,
    finalizer: 595
  },
  totalTokens: 3648
}

Proxy

Pass a proxy URL using proxyUrl option:

const swarm = new Swarm(apiKey, {
  proxyUrl: 'http://127.0.0.1:8080'
});

This SDK is part of Neurounit AI — my personal platform that helps businesses leverage artificial intelligence for automated lead generation.
I develop this project in my spare time.

💜 Support

If you'd like to support this project, you can:

• Donate via TRC20: TBJWkKyrQEmG1dr8rd3psDwWY4tEvUFnzw
• Mention it on social media 🙌
• Star ⭐ the repo to show appreciation