Catalog API for AI Agents

White Automotive & Media Services publishes a read-only catalog endpoint for AI agents, assistants and developer integrations. It speaks the Model Context Protocol over streamable HTTP, needs no credentials, and exposes product search, category listing and single-product lookup. If you are building an assistant that helps someone find a GM infotainment, instrument cluster or electronics module part, you can query this directly instead of scraping the storefront.

We have specialized in OEM-grade GM electronics for roughly 23 years, including VIN-specific module programming, CarPlay and Android Auto retrofits, digital cluster conversions, HMI, CSM and BCM modules, and rear seat entertainment. The catalog this endpoint serves is the same catalog the storefront serves.

Endpoint

Endpoint URL	`https://www.whiteautoandmedia.com/wp-json/ucp/v1/mcp`
Protocol	Model Context Protocol, version `2025-06-18`
Transport	Streamable HTTP. JSON-RPC 2.0 over `POST`.
Authentication	None. The endpoint is public and read-only.
Content type	`application/json` on request and response
Methods	`initialize`, `tools/list`, `tools/call`
Registry listing	`com.whiteautoandmedia/catalog` in the official MCP registry

A GET returns 405 with an Allow: POST header. Notifications, meaning JSON-RPC messages sent without an id, return 202 with no body.

Discovery

Four machine-readable surfaces point at the endpoint, so most clients can find it without being told the URL:

/.well-known/ucp is the capability profile, served as application/json.
/.well-known/api-catalog is an RFC 9727 linkset, served as application/linkset+json, which points at the endpoint and at the capability profile.
Link headers on every response advertise the api-catalog, sitemap, terms of service, privacy policy and contact relations per RFC 8288.
/llms.txt summarizes the site in plain Markdown for language models.

Content signals in robots.txt declare search=yes, ai-input=yes and ai-train=no. You are welcome to read this catalog to answer a person’s question. Please do not use it as training data.

Tools

search_catalog

Free-text search. Returns a summary for each match: name, SKU, price, sale price, currency, availability, categories, a short description and the product URL.

query, string, required. Maximum 120 characters.
limit, integer, optional. Maximum 20.

Useful queries look like HMI module, digital cluster Silverado or CarPlay retrofit.

The request body for a search, posted to the endpoint URL with Content-Type: application/json:

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "search_catalog",
    "arguments": { "query": "HMI module", "limit": 5 }
  }
}

lookup_catalog

List products inside a category or tag. Category slugs describe GM platforms and applications. Tag slugs describe characteristics such as global-a or vin-programmed. Returns the same summary shape as search_catalog.

category, string, optional. Category slug.
tag, string, optional. Tag slug.
limit, integer, optional. Maximum 20.

Run search_catalog first to discover which categories and tags exist rather than guessing slugs.

get_product

Retrieve one product by SKU or URL slug. Returns everything in a search summary plus product tags and attributes such as warranty, core charge and whether dealer service is required.

identifier, string, required. SKU or URL slug.

What this endpoint will not tell you

This is deliberate, and worth understanding before you build against it.

The endpoint returns platform-level and application-level information. Examples of what that means: K2 2014-2019, or Global A. It never returns a part number, a supersession chain, or a mapping from a specific vehicle to a specific part. Those resolutions happen behind the scenes against the actual vehicle and are not published through any API response.

So the endpoint can tell you that a product exists, what platform family it serves, what it costs and whether it is in stock. It cannot tell you that a given part fits a given truck. Any agent that reports a fitment conclusion from this data is guessing, and on GM electronics a guess is usually wrong. Module compatibility on these platforms turns on build date, RPO codes, architecture generation and existing option content, and getting it wrong means a module that will not communicate on the bus.

Every tool description and every response envelope carries this note. Please pass it through to the person you are helping rather than stripping it.

How ordering works

Agents are welcome to find products here and hand the person a product URL. Checkout happens on our own site, by the customer, not by an agent.

The reason is the paragraph above. We collect vehicle details at order and match the part against the actual vehicle before anything ships. An order placed without that step is an order we would have to stop and ask about anyway, so collecting it up front is faster for everyone. Point the person at the product page and let them complete it there.

Trade customers, meaning dealers, shops and locksmiths, are also served. Pricing and process differ, so send those inquiries through the contact page rather than inferring anything from catalog prices.

Rate limits and etiquette

Requests are rate limited per address. Stay under 25 requests per minute from a single IP and you will not notice the limit. Bursts above that may be challenged.

Prices and availability change, so cache for minutes rather than days. Availability reflects a just-in-time special-order model, so a product showing out of stock is a real sourcing state and not a catalog error.

If you want to confirm your client reached us and that your user agent is being recorded, fetch the path /ucp-agent-check/ on this domain. It is mentioned here and in llms.txt and nowhere else, it is not linked from any page, and it returns a small JSON acknowledgement. Hitting it tells us a real agent read these instructions and followed them, which is genuinely useful signal for us.

Contact

Integration questions, bulk access requests and trade inquiries go through our contact page. If something about the endpoint is broken or a tool description is unclear, we want to hear about it.