• This thread is just the tip of the iceberg.The people ahead of the curve aren't Googling for answers — they're already in here, having the conversations you haven't found yet. DealerRefresh is free.Get the full picture →

MCP from CDK & Reynolds and other DMS companies

What could possibly go wrong—especially inside a multimillion-dollar operation? ...

A lot.

It will come down to whether MCP connections can securely control exactly what an agent can see, what it can do, and when human approval is required. That said, based on what I’ve seen, I believe we are much closer than most people realize.
 
Yes, "WHAT could possibly go wrong giving an AI agent direct access to customer PII, driver's License, bank details, and full finance applications?"

That was my question when I was designing for this inside our system.

It turns out the solution is the same way we hide the information in transit and at rest

Browser side render of information that is token based linked to the authenticated user session.
 
Good thread. Alex, I’m not convinced CDK and Reynolds will both ship something with “MCP” on the box inside 12 months. I think a few vendors will do something, and I’ve already seen some examples. Either way, this is about giving dealers more power to use their own data. Using it well at the store level will take a mindset and skill shift.

If you’re on CDK, you don’t have to wait for it. I’ve built this myself at my two stores using CDK’s Data Your Way Import/Export API. It already gives you record-level read and writeback across Service, Vehicle, Customer, Parts, and Accounting. I’ve built an MCP server and a CLI on top of that same data access. In practice, I just point my Claude or OpenAI session at the CLI, and it uses the tools to answer questions across the data sets.

To Todd’s point: we’re working from our DMS, the system of record where the deals, ROs, and accounting actually get closed every month, so the data is pretty clean. The next step is building our own internal apps, which become another source of context for the agent.
 
Great discussion here. I think the bigger question isn't when CDK, Reynolds, and others will launch MCP servers, but whether they'll actually open up access to the data and capabilities dealers need to innovate.

We're building AI infrastructure for dealerships at CarGenius AI, including AI-powered service scheduling, and I can tell you firsthand how painful the current ecosystem is.

To build something as seemingly straightforward as an AI agent that books a service appointment, we have to navigate API certification fees, annual access fees, read fees, write fees, per-dealer fees, per-call fees, and more. Often across multiple providers just to deliver one experience.

It's insane. And dealers wonder why software is so expensive!

Imagine a different world:
- Need a new inventory feed? Grab it from a secure, documented endpoint.
- Want to build a new feature on your dealership website? Connect to the MCP.
- Want to build your own AI agent to answer customer questions, schedule service, or help your team operate more efficiently? Connect it to the systems you already use, with the appropriate permissions.

No months of integration work, expensive certification programs, or another toll gate at every step. That's one of the opportunities MCP represents. Not just another protocol, but a fundamentally more open ecosystem where dealers and developers can build and innovate.

Of course, security, permissions, and protecting customer data are critical. But we shouldn't confuse protecting data with protecting business models built around restricting access to it.

Dealers should be able to securely access their own data, authorize trusted partners, and choose the tools and experiences that work best for their business. The entire ecosystem benefits from more competition, faster innovation, and lower costs.

If MCP just becomes another line item on the DMS invoice, we've missed the opportunity entirely.
 
It will probably be a neutered list of tools and capabilities, but they'll be able to say they have an MCP Server for AI. Not just a them problem though, a lot of vendors that have an MCP end-point, have made it pretty vanilla out of the gate.

Some of the SAAS we use does, and it's great.

We have a full (read-only) MCP server connected to GA4, GAds, Search Console, Meta Ads, etc for all our managed dealers - have at it with your ChatGPT, Claude analylsis.
 
Will the automotive industry accept not having full read-write API access to their system of record?

So far we have. It has been to the detriment of local innovation. I think AI might be the equalizer at the dealership level. I could see why a single point system of record is important, but having as many bolt-on solutions that can't write back to it and take action will not be helpful.
 
What could possibly go wrong—especially inside a multimillion-dollar operation? ...

A lot.

It will come down to whether MCP connections can securely control exactly what an agent can see, what it can do, and when human approval is required. That said, based on what I’ve seen, I believe we are much closer than most people realize.

Agree, and I think the answer is as much architectural as it is about permissions.

A DMS should stay what it is: the system of record. The more we push MCP connections and agent logic into it, the more we risk slowing down a platform that dealerships rely on every single day. Stability there matters more than any new feature.

I'd keep the DMS as a reference point and put the AI agents outside it. The agents read the data, make the decisions, and take the actions, with human approval wherever the stakes require it. That way the DMS isn't loaded with extra work, and the control layer (what an agent can see, what it can do, and when a person signs off) lives in one place.

Separating record from decision keeps the core stable and lets the intelligence layer move at its own speed.
 

✨ AI Highlights

The thread debates when CDK, Reynolds, and other DMS providers will offer MCP (Model Context Protocol) connections to dealers, with participants skeptical that early releases will provide meaningful data access rather than just marketing-friendly labels. Key concerns include layered API fees, limited field exposure, read-only restrictions, and serious security risks around PII and financial data — with one dealer sharing that he built his own MCP server on CDK's existing Data Your Way API rather than waiting. The emerging consensus is that branded MCP products will arrive within 12 months but substantive, permissioned read-write access that enables real AI agent workflows could take two to three years unless competitive pressure, potentially from Tekion's upcharge model, forces faster movement.

Replies Views 16 1,367 Started Last Reply