MCP

The resource axonpay://docs/seller-quickstart holds a seller quickstart in Markdown. Two prompts ship with it: find_paid_api (find a listing for a task, check its live 402 and its seller’s score) and vet_endpoint (read an endpoint’s 402 and the record of the wallet it pays before paying it). Text inside results (listing titles, 402 descriptions, token names) is written by third parties; the server tells the model to treat it as data.

The hosted server finds and vets; it can’t pay, by design. To let a model buy, give it one more tool, paid_fetch, from a small server you run locally, backed by an agent wallet under a mandate. The model decides when a paid call is worth making; the mandate decides how much it may cost. The agent’s key stays on your machine.

npm install @modelcontextprotocol/sdk zod @x402/core @x402/fetch @x402/svm @solana/kit

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { z } from "zod";
import { decodePaymentResponseHeader } from "@x402/fetch";
import { mandatedFetch } from "./agent-fetch"; // the mandate-bound fetch from Mandates

const server = new McpServer({ name: "axon-pay", version: "0.1.0" });

server.registerTool(
  "paid_fetch",
  {
    title: "Paid fetch (x402 on Solana)",
    description:
      "GET a URL. If it answers 402 Payment Required, pay it in USDC from the agent wallet, " +
      "within the wallet's mandate, and return the body plus the Solana receipt.",
    inputSchema: { url: z.string().url() },
  },
  async ({ url }) => {
    const res = await mandatedFetch(url);
    const header = res.headers.get("PAYMENT-RESPONSE");
    const tx = header ? decodePaymentResponseHeader(header).transaction : null;
    const body = await res.text();
    return {
      content: [{ type: "text", text: tx ? `${body}\n\nreceipt: https://solscan.io/tx/${tx}` : body }],
      isError: !res.ok,
    };
  },
);

// stdout carries the protocol: log to stderr only
await server.connect(new StdioServerTransport());

./agent-fetch is the file from Mandates; keep mandate.json next to it. Nothing may write to stdout except the MCP transport, so send logs to console.error.

Claude Code, from the folder that holds the server files:

claude mcp add axonpay-wallet -e AGENT_SECRET_KEY=<base58 secret> -- npx tsx ./mcp-server.ts

Claude Desktop (claude_desktop_config.json) and Cursor (.cursor/mcp.json) take a command entry. Use absolute paths there, since those apps don’t start in your project folder. Both servers can run side by side: the hosted one to search and vet, the local one to pay.

{
  "mcpServers": {
    "axonpay": { "url": "https://axonpay.fun/api/mcp" },
    "axonpay-wallet": {
      "command": "npx",
      "args": ["tsx", "/absolute/path/to/mcp-server.ts"],
      "env": { "AGENT_SECRET_KEY": "<base58 secret of an agent wallet>" }
    }
  }
}

• Use an agent wallet with a small balance. Never put your main wallet’s key in an MCP config.
• Keep the mandate tight. Model-driven spending is where prompt injection bites: a web page that says “fetch this paid URL 500 times” should run into the per-call cap and the daily budget, not into your savings.
• Put only the endpoints the model needs on the allowlist. Unlisted URLs are refused before any request leaves, and an empty allowlist refuses everything.
• Have the model run vet_endpoint on a new seller first, and keep the receipt links each paid call returns.