partscout #

mcp server for pricing computer hardware on ebay — built with FastMCP for TypeScript.

written to answer one question honestly: should i buy this used?

why #

shopping for homelab gear means bouncing between listings trying to work out whether $300 for a used rack is a deal or a ripoff. that is a distribution question, not a lookup — so these tools return distributions.

tools #

caveats worth knowing #

• these are asking prices, not sold prices. the ebay Browse API exposes active listings only; sold data lives behind the restricted Marketplace Insights API. title filtering removes common false matches, but inspect candidate details before relying on either the minimum or median.
• GPU searches check titles, not just categories. RTX model/variant mismatches, empty boxes, loose coolers, waterblocks, NVLink bridges, missing-core cards, damaged/untested cards, and mobile eGPUs are excluded from desktop GPU searches. strict_gpu_model, include_accessories, and include_for_parts allow broader searches. This is a title heuristic, not a guarantee of condition or authenticity. Filtering scans at most five pages to fill the requested result limit.
• search results can be stale. Call get_listing with a result's item_id before recommending it. It refreshes the eBay item API's availability and price; unknown means unverified, not in stock. Shipping quotes may lack destination context; null is unknown, not free. Seller feedback count accompanies percentage.
• snapshots use USD fixed-price listings only, excluding shipping and tax. They describe the filtered best-match sample, not the entire market, past prices, warranty eligibility, or a recommendation to buy. Open-box remains in “new.”
• searches are category-scoped by default. keyword search alone is not enough: "Synology DS923+" unscoped returns ram upgrades and power bricks and prices them as if they were the NAS. every search is scoped to ebay's suggested category; pass auto_category: false to opt out, or pin category_id when the guess is wrong. suggest_categories shows the guesses.
• ebay's condition buckets are broader than they sound. NEW also returns "Open box" (kept — a real listing at a real price). USED also returns "For parts or not working" (dropped by default — a dead card is not a data point about what a working one costs). refurbished is in neither bucket.
• ebay often asks above retail for current-production goods. it is the right tool for used and discontinued hardware, the wrong one for anything you can still buy new from a retailer.

setup #

requires Node.js 22 or newer. register an application at developer.ebay.com and put the keys in the environment or a local .env:

EBAY_CLIENT_ID=your-app-id
EBAY_CLIENT_SECRET=your-cert-id
# optional
EBAY_MARKETPLACE_ID=EBAY_US
EBAY_BASE_URL=https://api.ebay.com   # or https://api.sandbox.ebay.com

the Browse API uses an application token (client credentials), so there is no user login step — it reads public listings and never acts on an account.

run #

npm install
npm run build
npm start                         # stdio, from the built package
npm run dev                       # interactive FastMCP development client

register the source checkout with Claude Code:

claude mcp add partscout -- npx --yes tsx /path/to/partscout/src/server.ts

or install the package and use its partscout-mcp executable.

use as a library #

import { Ebay } from "partscout";

const ebay = new Ebay();
const result = await ebay.newVsUsed("Synology DS925+");
console.log(result.verdict);

development #

npm run check

unit tests mock the ebay transport, so they need no credentials. live regression tests run automatically when EBAY_CLIENT_ID and EBAY_CLIENT_SECRET are set.