Why MCP matters for car dealers

In my last post, I introduced CarGenius and the bet behind it: car shopping is moving into AI, and dealers need a way to be there directly instead of getting disintermediated again.

This post is about the technology we are using to do it, MCP, which stands for Model Context Protocol.

MCP is a tool created by Anthropic, the makers of Claude, to help connect AI to external systems. It gives AI agents a structured way to use outside data and tools instead of relying on scraping, guessing, or one-off integrations.

That sounds technical, but the idea is simple:

Your website is built for humans. MCP is built for AI agents.

This matters because of where shoppers are going. A shopper who used to open Cars.com to figure out the best third-row SUV for their family now opens ChatGPT. Or Claude. Or Gemini. It answers the question, and it does it well.

The question for dealers is simple:

When that shopper's AI goes looking for an actual car, where does it get the data?

That is what MCP is about.

What MCP does

Think of MCP as an AI-native API.

An MCP server tells an AI agent what tools are available and how to use them. The agent can connect and ask, "What can you do?" The server might answer, "I can search inventory. I can return details on a specific VIN. I can explain the dealership's policies. I can submit a lead."

Then the AI can use those tools.

A simple example: imagine an MCP server for your Gmail. The AI asks what it can do. The server says, "I can read messages, check your calendar, and create invites." You ask, "What is on my calendar today?" and it reads it back. You ask it to create an invite, and it does that too.

You are talking in plain language, but the answer is coming from real systems.

That last part matters.

MCP gives AI tools, not just text. The model can reason in natural language, but the actual inventory results, pricing, VDP links, dealer policies, and leads come from your systems through the MCP server.

We are not training a model and hoping it remembers your dealership correctly.

We are giving the model tools that return your real data.

Why dealers should care

Right now, when a shopper asks an AI to help them find a car, the AI does the only thing it can do.

It scrapes.

It pulls from search results, marketplace pages, old listings, review sites, dealer websites, and whatever else it can find. Sometimes it does a decent job. Often it misses the most important parts.

It does not know you spent $2,000 reconditioning that two-year-old Accord. It may not have the Carfax. It may not understand your buying process, your fees, your service department, your trade process, your language capabilities, or why a shopper should work with you instead of the store down the street.

The marketplace never carried most of that information, so the AI cannot find it there.

Your website may have some of it. Your inventory feed may have some of it. Your CRM, service scheduler, trade tool, finance tool, and internal process may have even more of it.

MCP is how the AI gets to that information directly.

Instead of scraping a stale listing page, an agent can ask your MCP server:

"Show me used AWD SUVs under $30,000."

"Tell me everything about this VIN."

"Does this dealer work with out-of-state buyers?"

And the answer comes from you.

The starting point is changing. More consumers are beginning their research inside AI: figuring out what to buy, what to avoid, what is a fair deal, and what fits their life. The next step is obvious. They will ask AI to help them find the actual car.

If a dealer does nothing, AI will still shop. It will just use whatever it can find: incomplete listings, old reviews, generic website copy, missing fees, and marketplace data that may be two steps removed from the dealer.

With MCP, the dealer gets to control the source of truth. You get to hand AI better data than the marketplace ever had.

And that matters because AI is not going to show 120 results. It is going to recommend a handful of vehicles and explain why.

The dealers who win will be the ones whose inventory, story, policies, and next steps are easiest for AI to understand and act on.

What CarGenius does with MCP

We build the MCP server for the dealer.

It lives on the dealer's own domain, something like mcp.dealership.com, and exposes a set of tools for AI agents to use.

Today it does four things: it searches the dealer's live inventory in plain language, returns VIN-level detail with pricing, photos, availability, and a direct link back to the dealer's VDP, explains how the store actually operates from policies and fees to service and languages, and helps the shopper take the next step and reach the dealer directly through ADF or CRM.

That is the start, not the finish.

We already have Carfax integration live, so the vehicle history can travel with the vehicle data where the dealer has the right commercial relationship in place. That is something a marketplace usually cannot expose the same way.

Next come service scheduling, sales scheduling, trade-in, finance, and more.

This is the bigger point: MCP starts as discovery infrastructure, but it becomes the transaction layer for AI-assisted car buying. AI should not just be able to tell a shopper that a car exists. It should help them move forward.

Where MCP discovery stands today

I want to be direct about the current state, because the real version is better than the hyped version.

ChatGPT is not automatically using every dealer's MCP server today.

That is the reality.

Here is what happens now. We deploy the MCP server on the dealer's domain and publish AI-facing discovery signals, including things like an llms.txt file and an AI resources page. Those signals help AI systems understand that the dealership has an MCP server and what tools are available.

In our experience, an AI like ChatGPT can already find that the server exists and understand what it is meant to do. What it generally will not do yet is execute that server on behalf of a shopper. For now, it is boxed in by its own environment and blocked from calling the tools unless the user explicitly adds the MCP server themselves.

Here is what we tell dealers today:

ChatGPT will not start shopping your MCP server tomorrow. But it will know you have it, and you are getting ready for the moment when agents can use it directly.

That is still valuable. In fact, it is exactly when dealers should be doing this work. The time to prepare your AI-facing infrastructure is before AI becomes a major shopping channel, not after.

Where MCP discovery is going

We think the future is becoming clear.

AI agents need a better way to discover and use trusted tools on the internet. Scraping is too messy. Custom integrations do not scale. Marketplaces are incomplete. And consumers are going to expect their AI to do more than summarize search results.

The future is agents discovering tools on their own.

For dealers, that means an AI agent should be able to discover that your dealership has an MCP server, understand what tools it offers, and decide when to use those tools during a real shopping conversation.

One emerging mechanism is a server card. Think of it as a small, AI-readable file on your domain that tells agents three things: this MCP server exists, here are the tools it offers, and here is how to use them.

It is a sitemap for AI agents.

There are two gates to get through. The first is discoverability: ChatGPT, Claude, Gemini, and other agents need to know your server exists. The second is automatic usage: they need to be able to call that server when a shopper asks a relevant car-buying question.

We are building for both. Today, we can publish the server, expose the discovery signals, and use the MCP through AI Search on the dealer's site. Tomorrow, as agent discovery and usage mature, that same infrastructure becomes the way outside AI agents shop your store directly.

The dealers who build this infrastructure early will have a better shot at being in that answer.

Why AI Search matters now

All of that raises a fair question: if AI agents are not shopping your MCP server yet, why act today?

Because you do not have to wait for them to get value from this.

When we set up a dealer's MCP server, we also pair it with a leading AI model and put it on the dealer's website. We call that AI Search. It is essentially ChatGPT on your site, powered by your inventory, your policies, your website, and your tools.

Dealers have told us it feels like their dream salesperson. It knows the inventory. It knows the store. It runs 24/7. It does not forget anything. And it lets the customer shop the way they already want to shop.

But here is the bigger point. AI Search is not separate from MCP. It is the first live application of the dealer's MCP server. The same tools that power the on-site experience are the tools outside AI agents will use later. So every time we improve AI Search, we are improving the dealer's MCP.

It also becomes a rehearsal for the future. You get to see the questions shoppers actually ask, and how the AI describes your store. When the AI gives a weak answer about your buying process, we find the gap and fix it. When your website tells an inconsistent story or your sitemap is missing important pages, the AI exposes it and we catch it. When the vehicle data is thin, we add depth.

That makes the consumer experience better today. And it makes your MCP better for the moment outside AI agents come shopping tomorrow.

Build the rails now

This is a flywheel. As more dealers stand up MCP servers, AI gets better data. As AI gets better data, the shopping experience gets better. As the experience gets better, shoppers use AI more. And as shoppers use AI more, dealers need to be ready.

We are not building our own consumer marketplace. We are not trying to replace the dealer. We are helping dealers show up directly where shoppers are going.

Getting started is simple. We can stand up your MCP server, connect your inventory, publish the AI-facing discovery layer, and turn on AI Search so you can start seeing how AI understands your store.

Let's get your dealership ready.