MCP server development

MCP servers for dealers, built for AI assistants.

Creative Code builds MCP (Model Context Protocol) servers for car dealers and automotive software companies, so AI assistants such as Claude and ChatGPT can query live inventory, offers and store data through tools instead of scraping a website. We run dealer MCP servers in production today, including an inventory server with 19 public read-only tools and an agent storefront with 18 dealer tenants live.

Talk about an MCP serverSee the storefront

Definition

What is a dealer MCP server?

The Model Context Protocol (MCP) is the open standard AI assistants use to call external tools and read external data. A dealer MCP server exposes dealership data and actions as tools: search inventory, get a vehicle, list current offers, get store hours, or send a question to the store.

When a shopper asks Claude or ChatGPT about a car, the assistant calls those tools and answers from live data, instead of guessing from a crawl of the dealer website that may be days old.

Tools

Tool categories we build

Inventory

Search by make, model, price and location; vehicle detail by VIN; new arrivals, aging and recently sold units.

Pricing and offers

Price evaluation, price drops, OEM incentives matched to vehicles, and finance, lease and trade math.

Market data

Market statistics, model comparisons and dealer rankings from the inventory graph.

Store information

Hours, location and store FAQs, so the assistant answers in the dealer's own words.

Leads and appointments

Send a question or lead to the store, schedule a test drive or service appointment. Write tools sit behind a consent step.

Group analytics

Metric queries for dealer groups, so an AI analyst can answer with the numbers attached and the definitions behind them.

Security

Auth, tenant isolation and consent gates

Authentication

Public tools are read-only and need no sign-in. Account tools need an API key, and interactive clients sign in with OAuth. Write tools never ride on the public tier.

Tenant isolation

Each dealer or group is its own tenant with its own credentials, encrypted at rest. Every tool call is scoped to one tenant.

Consent gates

Any tool that reaches a person, such as sending a lead or scheduling an appointment, requires the shopper to approve first. Nothing is sent without that record.

Transport and clients

MCP over Streamable HTTP. We test with Claude custom connectors and ChatGPT apps, and other MCP clients such as Cursor connect the same way.

Open standard

A protocol, not a one-off

We published a draft open specification, the Open Protocol for Dealership AI Agents (RFC 0.1). It defines eight dealership domains over MCP, so an agent can reach inventory, CRM, desking, service and parts without a bespoke integration for each vendor pair.

How an engagement works

Discover, deploy, measure

Discover

01

Weeks 1 to 4

A Forward Deployed Engineer maps the real process in your systems, measures the baseline, proves the hard parts with prototypes, and delivers an architecture and a scope you can take to anyone, including us.

Deploy

02

Weeks 5 to 8

The engineering pod and the Farm, our AI agent fleet, build to production inside your systems of record. Every change ships with tests and evidence, and an engineer merges.

Measure

03

Months 3 and 6

Results against the baseline, in the same units we measured on day one. Then we keep improving what is live.

Dealer MCP server FAQ

What is an MCP server for a dealership?+

A server that speaks the Model Context Protocol, the open standard AI assistants use to call external tools. For a dealership it exposes tools such as search inventory, get a vehicle, list offers or get store hours, so an assistant like Claude or ChatGPT answers from live dealer data instead of a stale crawl of the website.

Do you have dealer MCP servers in production?+

Yes. Our inventory MCP server exposes 19 public read-only tools (inventory search, vehicle detail, pricing, offers, market and dealer statistics) plus account tools behind an API key. Our AI Agent Storefront serves dealer inventory and store data over MCP for 18 tenants.

Which AI assistants can connect?+

Any client that supports MCP over Streamable HTTP. We test against Claude (custom connectors) and ChatGPT (apps), and other MCP clients such as Cursor connect the same way.

Can the server create leads or schedule appointments?+

It can, as write tools scoped per project. Any tool that reaches a person, such as sending a lead to the store or scheduling a test drive or service appointment, sits behind a consent step the shopper completes first. Nothing is sent without it.

How is dealer and shopper data kept separate?+

Each dealer or group is its own tenant with its own credentials, encrypted at rest, and every tool call is scoped to that tenant. Public tools are read-only. Account tools need a key, and interactive clients sign in with OAuth.

Does it replace the dealer website?+

No. The dealer keeps the website they have. The MCP server is a second front door built for AI assistants, and it can report which assistants read the store and what they asked for.

Is there a standard for dealership MCP tools?+

We published a draft open specification, the Open Protocol for Dealership AI Agents (RFC 0.1), which defines eight dealership domains over MCP so any agent can reach inventory, CRM, desking, service and parts without bespoke work.

How long does an MCP server take to build?+

Discover runs four weeks and proves the hard parts with a working prototype. Deploy follows in weeks five to eight. Projects that sit on our existing inventory data and connectors skip most of the data work.

Scope an MCP server

Related