scaffold-agent

You are scaffolding a new three.ws agent into the user's project. The argument $ARGUMENTS is the agent name (default: MyAgent if blank).

Step 1 — Detect project type

Check the current working directory for:

• package.json → Node.js / browser project
• vite.config.* → Vite frontend
• next.config.* → Next.js
• No package.json → standalone HTML

Read package.json if present to understand the existing stack.

Step 2 — Install dependencies

For Node.js projects, run:

npm install @three-ws/sdk @x402/mcp @modelcontextprotocol/sdk/client

If the project uses a lockfile other than package-lock.json (yarn.lock → yarn, pnpm-lock.yaml → pnpm), use the right package manager.

Step 3 — Create the agent entry file

Create agent.js (or agent.ts if TypeScript is detected) in the project root with this structure, substituting $ARGUMENTS for the agent name:

import { AgentKit } from '@three-ws/sdk';
import { Client } from '@modelcontextprotocol/sdk/client/index.js';
import { StdioClientTransport } from '@modelcontextprotocol/sdk/client/stdio.js';

// Connect to the three.ws paid MCP tools
const transport = new StdioClientTransport({
  command: 'npx',
  args: ['-y', '@three-ws/mcp-server'],
  env: {
    MCP_EVM_PAYMENT_ADDRESS: process.env.EVM_WALLET_ADDRESS ?? '',
    MCP_SVM_PAYMENT_ADDRESS: process.env.SOLANA_WALLET_ADDRESS ?? '',
  },
});

const mcp = new Client({ name: '<AgentName>', version: '1.0.0' });
await mcp.connect(transport);

// Mount the three.ws chat + avatar panel
const agent = new AgentKit({
  name: '<AgentName>',
  description: 'A three.ws AI agent',
  endpoint: process.env.AGENT_ENDPOINT ?? 'http://localhost:3000',
  onMessage: async (text, context) => {
    // Your LLM call goes here — replace with Anthropic, OpenAI, or Vercel AI SDK
    return `Echo: ${text}`;
  },
});

// Use paid MCP tools
export async function getPose(prompt) {
  const result = await mcp.callTool('get_pose_seed', { prompt });
  return JSON.parse(result.content[0].text);
}

export async function getTokenSnapshot(mint) {
  const result = await mcp.callTool('pump_snapshot', { mint });
  return JSON.parse(result.content[0].text);
}

export { agent, mcp };