Build an Agent with aixyz

When to Use

Use this skill when:

• Scaffolding a new AI agent project from scratch
• Adding a new tool to an existing agent
• Configuring x402 micropayments for an agent or tool
• Wiring up A2A and MCP protocol endpoints
• Deploying an agent to Vercel

Instructions

1. Scaffold a new project

All CLI commands support --help for full usage details. Use --help to discover available options.

# See all options
bunx create-aixyz-app --help

# Interactive (TTY)
bunx create-aixyz-app my-agent

# Non-interactive (recommended for AI/CI)
bunx create-aixyz-app my-agent --yes
bunx create-aixyz-app my-agent --erc-8004 --pay-to 0x... --no-install

--openai-api-key is optional — if omitted, set OPENAI_API_KEY in .env.local before running the agent. The scaffolded template uses @ai-sdk/openai by default, but you can swap it for any Vercel AI SDK provider adapter (e.g. @ai-sdk/anthropic, @ai-sdk/google, @ai-sdk/amazon-bedrock).

This creates the standard project layout:

my-agent/
  aixyz.config.ts     # Agent metadata and skills
  app/
    agent.ts          # Agent definition
    tools/            # One file per tool
    icon.png          # Agent icon (optional)
  package.json
  vercel.json

2. Configure the agent (aixyz.config.ts)

Every agent needs a config file at the project root. Declare identity, payment address, and skills:

import type { AixyzConfig } from "aixyz/config";

const config: AixyzConfig = {
  name: "My Agent",
  description: "A short description of what this agent does.",
  version: "0.1.0",
  x402: {
    payTo: process.env.X402_PAY_TO!,
    network: process.env.NODE_ENV === "production" ? "eip155:8453" : "eip155:84532",
  },
  skills: [
    {
      id: "my-skill",
      name: "My Skill",
      description: "What this skill does for callers.",
      tags: ["example"],
      examples: ["Do something with my skill"],
    },
  ],
};

export default config;

3. Payment: accepts export (aixyz/accepts)

Every agent and tool controls whether it requires payment by exporting accepts from aixyz/accepts. Without this export, the endpoint is not registered for payment gating.

import type { Accepts } from "aixyz/accepts";

// Require x402 micropayment
export const accepts: Accepts = {
  scheme: "exact",
  price: "$0.001", // USD-denominated
  network: "eip155:8453", // optional — defaults to config.x402.network
  payTo: "0x...", // optional — defaults to config.x402.payTo
};

// Or make the endpoint free
export const accepts: Accepts = {
  scheme: "free",
};

Export accepts from app/agent.ts to gate the A2A endpoint, or from a tool file to gate that tool via MCP.

4. Write a tool (app/tools/<name>.ts)

Each file in app/tools/ exports a Vercel AI SDK tool as its default export:

import { tool } from "ai";
import { z } from "zod";
import type { Accepts } from "aixyz/accepts";

export const accepts: Accepts = { scheme: "exact", price: "$0.001" };

export default tool({
  description: "A short description of what this tool does.",
  inputSchema: z.object({
    query: z.string().describe("Input to the tool"),
  }),
  execute: async ({ query }) => {
    // your logic here
    return { result: query };
  },
});

Files prefixed with _ (e.g. _helpers.ts) are ignored by the auto-generated server.

5. Define the agent (app/agent.ts)

The default template uses @ai-sdk/openai, but you can use any Vercel AI SDK provider adapter:

// OpenAI (default in template)
import { openai } from "@ai-sdk/openai";

// Or swap the model provider — install the adapter and change the import:
// import { anthropic } from "@ai-sdk/anthropic";  // set ANTHROPIC_API_KEY
// import { google } from "@ai-sdk/google";        // set GOOGLE_GENERATIVE_AI_API_KEY

import { stepCountIs, ToolLoopAgent } from "ai";
import type { Accepts } from "aixyz/accepts";
import myTool from "./tools/my-tool";

export const accepts: Accepts = { scheme: "exact", price: "$0.005" };

export default new ToolLoopAgent({
  model: openai("gpt-4o-mini"), // replace with anthropic("claude-..."), google("gemini-..."), etc.
  instructions: "You are a helpful assistant.",
  tools: { myTool },
  stopWhen: stepCountIs(10),
});

5a. Add sub-agents (app/agents/<name>.ts)

Place additional agent files in app/agents/ to expose multiple independent A2A endpoints from one deployment. Each filename becomes a URL prefix:

app/
  agent.ts           # → /agent  (main)
  agents/
    math.ts          # → /math/agent
    text.ts          # → /text/agent

Each sub-agent file has the same structure as app/agent.ts:

import { openai } from "@ai-sdk/openai";
import { stepCountIs, ToolLoopAgent } from "ai";
import type { Accepts } from "aixyz/accepts";
import calculate from "../tools/calculate";

export const accepts: Accepts = { scheme: "exact", price: "$0.001" };

export default new ToolLoopAgent({
  model: openai("gpt-4o-mini"),
  instructions: "You are a math specialist. Use the calculate tool for arithmetic.",
  tools: { calculate },
  stopWhen: stepCountIs(5),
});

Sub-agents share the same /mcp endpoint for tools and each get their own agent card:

• /math/.well-known/agent-card.json
• /math/agent

See examples/agent-with-sub-agents for a working example with a coordinator + two specialists.

6. Environment variables (.env files)

Environment variables are loaded in the same priority order as Next.js:

1. .env.<NODE_ENV>.local (highest priority; not loaded when NODE_ENV=test)
2. .env.local
3. .env.<NODE_ENV> (e.g. .env.production, .env.development)
4. .env

Common variables:

7. Agent icon (app/icon.png)

Place an icon file at app/icon.png (also accepts .svg, .jpeg, .jpg). During aixyz build it is:

• Copied to the output as icon.png
• Converted to a favicon.ico (32×32) and placed in public/

No configuration needed — the build step auto-detects and processes the icon.

8. Custom facilitator (app/accepts.ts)

By default, aixyz uses https://x402.use-agently.com/facilitator to verify payments. To use a different facilitator, create app/accepts.ts and export a facilitator:

import { HTTPFacilitatorClient } from "aixyz/accepts";

export const facilitator = new HTTPFacilitatorClient({
  url: process.env.X402_FACILITATOR_URL ?? "https://www.x402.org/facilitator",
});

9. Run the dev server

bun run dev      # aixyz dev — starts at http://localhost:3000 with hot reload
bun run dev -- -p 4000  # custom port

Endpoints served automatically:

10. (Optional) Custom server (app/server.ts)

For full control, create app/server.ts. It takes precedence over auto-generation. The accepts field in mcp.register is optional — omit it to expose the tool without payment gating:

import { AixyzServer } from "aixyz/server";
import { useA2A } from "aixyz/server/adapters/a2a";
import { AixyzMCP } from "aixyz/server/adapters/mcp";
import * as agent from "./agent";
import myTool from "./tools/my-tool";

const server = new AixyzServer();
await server.initialize();
server.unstable_withIndexPage();

useA2A(server, agent);

const mcp = new AixyzMCP(server);
await mcp.register("myTool", {
  default: myTool,
  // accepts is optional — omit to expose without payment
  accepts: { scheme: "exact", price: "$0.001" },
});
await mcp.connect();

export default server;

11. Build and deploy to Vercel

bun run build    # aixyz build — outputs Vercel Build Output API v3 to .vercel/output/
vercel deploy

Examples

Working examples in the repo: examples/agent-boilerplate, examples/agent-price-oracle, examples/agent-byo-facilitator, examples/agent-with-sub-agents.

Common Dependencies

A scaffolded agent project uses these key packages:

To use a different LLM provider, install its AI SDK adapter and update the import in app/agent.ts:

bun add @ai-sdk/anthropic   # Anthropic Claude
bun add @ai-sdk/google      # Google Gemini
bun add @ai-sdk/amazon-bedrock  # AWS Bedrock

See the Vercel AI SDK providers for the full list.

Common Edge Cases

• Missing x402.network — always provide x402.network; it has no fallback.
• Missing x402.payTo — set X402_PAY_TO in .env.local or provide it directly in config.
• Tool file ignored — files prefixed with _ are excluded; rename to remove the prefix.
• Agent card missing skills — skills defaults to []; add at least one entry to be discoverable.
• Free endpoint — export accepts: { scheme: "free" } to expose an endpoint without payment.
• Port conflict in dev — use aixyz dev -p <port> to change the default port (3000).

CLI Reference (Non-TTY / AI-Friendly)

All CLI commands are designed for non-interactive use. When stdin is not a TTY, prompts are skipped — values come from CLI flags, environment variables, or sensible defaults. Use --help on any command for full usage.

create-aixyz-app

bunx create-aixyz-app --help

aixyz dev / aixyz build

aixyz dev --help
aixyz build --help

aixyz erc-8004 register

aixyz erc-8004 register --help

aixyz erc-8004 update

aixyz erc-8004 update --help