Give Claude or ChatGPT Real eBay Sold Prices (Tool Calling and MCP)

Ask an AI assistant what a used Nintendo 64 sells for and you'll get a confident number. Ask where it came from and things get vague. eBay blocks most AI crawlers in its robots.txt and hides sold results behind a sign-in, so assistants are working from old or secondhand data.

The fix is a tool: give the agent a function that fetches real sold prices.

#The failure to design against

One open-source eBay tool's issue tracker describes it well: the assistant "didn't surface the failed tool call and instead searched the web for completed listings," then answered anyway. The user got a number with no real data behind it.

So the tool's job isn't only to return prices. It has to make "I don't know" impossible to paper over.

#The tool

// One tool, explicit statuses. Works for OpenAI, Anthropic or any tool-calling model.
export const soldPriceTool = {
  name: "get_sold_prices",
  description:
    "Recent eBay sold prices (last 90 days) for an exact item. Use precise queries: brand + model, " +
    "or card name + set + number + grade in quotes. Quote only the numbers returned. If status is not 'ok', say no reliable price is available.",
  input_schema: {
    type: "object",
    properties: {
      q: { type: "string", description: "e.g. 'nintendo 64 console' or 'charizard 4/102 \"psa 9\"'" },
      condition: { type: "string", enum: ["new", "used", "for_parts"] },
    },
    required: ["q"],
  },
};

export async function getSoldPrices({ q, condition }) {
  const url = new URL("https://api.soldcomps.io/v1/sold");
  url.searchParams.set("q", q);
  if (condition) url.searchParams.set("condition", condition);
  const res = await fetch(url, { headers: { Authorization: `Bearer ${process.env.SALEHISTORY_KEY}` } });
  if (!res.ok) return { status: "unavailable", http: res.status };
  const d = await res.json();
  const sales = d.count - d.stats.excluded;
  if (sales === 0) return { status: "no_sales", query: q };
  return {
    status: sales < 5 ? "thin" : "ok",
    median: d.stats.median, p25: d.stats.p25, p75: d.stats.p75, sales, currency: d.stats.currency,
    recent: d.items.slice(0, 5).map((i) => ({ title: i.title, price: i.sold_price.value, sold_at: i.sold_at, url: i.url })),
  };
}

Returning recent sales with links lets the model cite them, and lets the user check them.

#As an MCP server

To use it from Claude Desktop or another MCP client, wrap the same function in a tiny MCP server:

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { z } from "zod";
import { getSoldPrices } from "./sold.js";

const server = new McpServer({ name: "ebay-sold-prices", version: "1.0.0" });
server.tool(
  "get_sold_prices",
  "Recent eBay sold prices for an exact item. Quote only returned numbers.",
  { q: z.string(), condition: z.enum(["new", "used", "for_parts"]).optional() },
  async (args) => ({ content: [{ type: "text", text: JSON.stringify(await getSoldPrices(args)) }] }),
);
await server.connect(new StdioServerTransport());

Register it in your client's MCP config with SALEHISTORY_KEY in its environment.

#System prompt rules that help

• Use the tool for any "what's it worth / what does it sell for" question.
• Build precise queries; for cards, one query per grade.
• Quote the median and the sale count; link a recent sale.
• If the status isn't ok, say there's no reliable price. Never estimate.

#Why agents need a hosted source

An agent that scrapes eBay itself runs into the same walls scrapers do: the sign-in wall, IP blocks, and pages that parse as "0 results." A hosted API turns that into one call that either returns data or a clear error. Why empty results are the dangerous failure.