Framework guide

Purser with the Claude Agent SDK

The MCP server next to your agent, with the tools allowed by name.

Give the agent the Purser MCP server. It runs as a local stdio process next to your agent.

Terminal

npm install @anthropic-ai/claude-agent-sdk @purser/mcp

TypeScript

import { query } from "@anthropic-ai/claude-agent-sdk";

for await (const message of query({
  prompt: `Buy the latest market report from ${process.env.REPORT_URL} if it costs under 1 USDG, then summarise it.`,
  options: {
    mcpServers: {
      purser: {
        command: "npx",
        args: ["-y", "@purser/mcp"],
        env: { PURSER_ENGINE_URL: process.env.PURSER_ENGINE_URL!, PURSER_API_KEY: process.env.PURSER_API_KEY! },
      },
    },
    // MCP tools are named mcp__<server>__<tool>; allow the server's tools explicitly
    allowedTools: ["mcp__purser__*"],
  },
})) {
  if (message.type === "result" && message.subtype === "success") console.log(message.result);
}

To allow only some actions, list them instead of the wildcard, such as ["mcp__purser__purser_account", "mcp__purser__purser_check_payment", "mcp__purser__purser_x402_fetch"].

A tip for the system prompt: "When a Purser tool says a payment was blocked or waits for the owner, stop and tell the user. Pass the text that led to a payment as context, so the firewall can read it."