Real World MCP: Home Depot

Home Depot doesn’t have a public MCP server yet, so this is a look at how one could turn AI conversations into sales.

When a homeowner asks ChatGPT, Claude or Gemini how to fix a rotted deck board, the sale is often decided before they ever open a browser tab. Most AEO work so far has been passive: structuring web content so LLMs cite your brand. An MCP server goes a step further and lets the AI check stock, build a cart, and send the customer to checkout. I covered the basics of this in MCP Discovery Servers for AI Agents. Home Depot makes a good real world example.

What a Home Depot MCP Server Provides

MCP (Model Context Protocol) is an open standard that lets AI assistants query outside systems and take actions in real time. Instead of relying on old training data or web scraping, an LLM connected to a Home Depot MCP server could ask Home Depot directly for:

• Local inventory at the customer’s preferred store or ZIP code
• Current pricing and promotions
• Exact product names, SKUs and specs from the product catalog
• Aisle and bay locations inside the store
• Product compatibility (which dimmer works with which LED fixture, which trim matches which siding)

Add account login through OAuth 2.1 and the assistant can act for the customer:

• Save a project list in their Home Depot account (or Pro Xtra account)
• Add every part needed for the job to their cart
• Give them a link for checkout, curbside pickup, delivery, or an in-store map

Example: Spotlight Bulb Replacement

Say a homeowner uploads a photo of their backyard to Claude and asks:

I’m replacing the lights on my backyard security spotlight and want to spread light more evenly across the yard. Here’s my current setup. If I want to replace just the bulbs, what should I get?

Without an MCP server, the AI gives generic advice (“look for a PAR38 flood bulb with a wide beam angle”), and the customer probably buys whatever shows up first on Amazon.

With a Home Depot MCP server connected, it could go like this:

1. Photo check: the LLM looks at the photo and identifies the fixture type and socket base.
2. Catalog lookup: it calls a tool like search_compatible_products, filtered to what’s in stock at the customer’s local store.
3. Specific recommendation: it suggests an exact PAR38 LED bulb with a wider beam angle and more lumens, with current local pricing for the number of bulbs needed.
4. Login: it offers to save the list or build the cart, and the customer logs in.
5. Checkout: the customer gets a checkout link, or a list showing the aisle, bay and total price.

Example: Rotted Siding Repair

My siding is rotted in this spot (see image). What products do I need to replace this, and are they in stock at my local store?

This one needs several products, not one: siding panels, J-channel trim, house wrap, corrosion resistant nails, exterior caulk, and maybe a pry bar.

When an LLM guesses at this list, the customer is left hunting down six items across different websites. An MCP server turns it into a single cart where every item is compatible, in stock at the same store, and ready for pickup. For a retailer, that’s where the real revenue is: complete project baskets, not single items.

How LLMs Handle This Today

Right now LLMs learn about retail catalogs through web crawls and live search. For a big retailer that leaves real gaps:

Capability	Web Crawls & Search	MCP Server
Inventory	Guessed or out of date	Live, store level stock
Pricing	Often weeks or months old	Current pricing and local promos
Products	Generic (“LED flood bulb”)	Exact SKUs, specs and compatibility
In-store location	None	Aisle and bay
Checkout	Customer leaves the chat to shop	Cart or list built in the chat
Competitors	AI may recommend another store	Your catalog is the answer
Attribution	Mostly invisible	Trackable tool calls and carts

An LLM will never keep a live copy of a huge retail catalog on its own. Store stock changes constantly. If you want AI assistants to recommend your products with confidence, you have to hand them structured, accurate data directly.

E-Commerce & AEO

An MCP server sits between two things that usually belong to different people:

• For the CMO (AEO and e-commerce): as product research moves from search engines to AI assistants, your brand becomes the source the AI relies on to answer the question, and every cart it builds can be attributed. I wrote more about tracking this in AI Attribution.
• For the CTO (controlled data access): instead of fighting AI scraper bots hitting your servers, you offer a rate limited endpoint that exposes only what you want LLMs to see.

Technically you’re setting up a separate system for people to buy your products, one that lives inside the AI tools they already use.

Caching Keeps Costs Low

The obvious question from engineering is whether millions of LLM tool calls would overload inventory and ERP systems. With the right setup the load is small:

1. Cache the static data heavily. Most of what the server returns (product descriptions, specs, compatibility, store layouts) rarely changes and can be served from a CDN or Redis cache.
2. Cache inventory for short windows. Price and stock change often, but caching store level availability for 5 to 15 minutes absorbs big spikes without hurting accuracy.
3. Separate read and write tools. Discovery tools (search_catalog, check_store_inventory) run without login at the edge. Cart tools (create_cart, add_to_pro_list) only run after OAuth login, so the expensive calls come from people ready to buy.
4. Reuse the existing APIs. An MCP server doesn’t need a new backend. It’s a thin layer on top of the REST or GraphQL APIs that already power the mobile app and website.

Compared to the sales it could drive, the cost to run it would be minor.

MCP Server Setup Help

If you’re the CTO or CMO at a company that could benefit from something like this, I may be able to help. Take a look at my AEO consulting work, or get in touch to talk about how an MCP server could work for your catalog.