AutomotiveMCP Brief #001: Shopify and Google open up to AI agents, and our test kit passes 9,900 calls

Hey, it's David, Brian, Gray, Daniel and Alex, the founding team behind AutomotiveMCP. Welcome to our very first Brief. We meant to keep issue one short. Then Shopify made every one of its stores reachable by AI agents, Google opened Gemini and Search to any company with an MCP server, and our own test kit got busy. So, not short. Grab a coffee. Here's what shipped and what to do about it.

New here? AutomotiveMCP is an open, industry-wide effort to make every dealership tool ready for AI agents. That means one shared front door (a gateway) so ChatGPT, Claude or your own agent can work with a store's tools, an open spec, a free test kit to try it, and a public directory of which dealer tools are ready. Vendors don't have to reshape their data or adopt one protocol: an MCP server or a secure, documented API both count. Why we're doing this · Get involved

Quick refresher: MCP (Model Context Protocol) is how an AI agent plugs into a tool and calls it. WebMCP is the same idea inside a web page, so an agent in the browser can call a site's tools instead of clicking buttons.

📣 Building something with agents at a dealership? Write to builders@automotivemcp.ai. We want to feature real dealer builds here, even small ones.

🏠 AutomotiveMCP update: 9,916 tool calls since the test kit went live

📊 The test kit

The test kit connects ChatGPT, Claude or your own agent to a sandbox store, Castle Kia of McHenry, IL, with simulated vendor tools behind one gateway. You can shop, see offers, price a trade, get a payment, approve sharing and pick a service time. A live session page shows every vendor call and consent receipt, with a Revoke button. On Oct 21 we posted a walkthrough of the test kit (the video is on that post).

Since it went live on Oct 19, AI agents and crawlers have made 9,916 tool calls across 4,404 agent sessions. 9,392 of those calls, across 4,266 sessions, came in the last 7 days. These numbers leave out our own internal test sessions.

The most-called tools:

Vehicle detail lookups: 1,746
Payment quotes: 1,523
Incentive and offer checks: 1,365
Trade-in appraisals: 533
Inventory searches: 235

Why care: Every one of those calls ran against the same sandbox a vendor's real tools would sit behind, through the same gateway, with the same consent and receipts. That's the path a real dealer integration takes.

Gateway (no sign-in needed):
https://gateway.automotivemcp.ai/mcp

Try the test kit · Watch the walkthrough

Welcome aboard. 15 organizations have applied to be founding members so far: 9 dealer groups, 3 vendors and 3 partners. Thank you, and welcome to everyone who's joined. We're glad you're here.

🤝 Halo is our first Founding Vendor Partner

Halo, the company behind the dealer-first AI operating system Halo OS, is opening its internal dealer business tooling to MCP so agents can take over the tedious work that eats a store's day. It's building to the open AutomotiveMCP spec as a Founding Vendor Partner, while Halo OS stays in closed beta with dealer groups ahead of a Q1 2027 public launch.

🧵 The builders' forum

forum.automotivemcp.ai is where vendors, dealers and in-store builders compare integration notes: how sign-in works, which tools are exposed, what's read-only, and what still needs a person to approve it. Every vendor in the directory gets its own thread. This week we added MCP setup notes for four more dealer tools:

AutoDealer.io: remote MCP on all plans at https://autodealer.io/api/mcp. Each connection only sees what that user's role allows, so a salesperson can't see the floor plan. It can't change anything by default. Turn on "prepare changes for my approval" and every change waits on an approval card until a person signs off.
DealerAI: a Zapier-hosted MCP with three tools that only write: add a contact, send an ADF, route a lead to the CRM. There are no read tools. Treat the endpoint URL like a password, and note that each call uses 2 Zapier tasks.
Chaiz: read-only service contract quotes at https://mcp.chaiz.com/mcp, with no API key. You can search by VIN, plate or make/model/year. The purchase happens on chaiz.com.
Vehiso: DMS MCP with OAuth or a secret key. It can read stock, enquiries, customers, appointments and deals, and make everyday stock changes such as price and status. Your DMS permissions still apply, so connect as a user whose access you'd trust an agent with.

MCP notes are also up for Fullpath, Hrizn, Space Auto, MarketCheck, CallRail, DealerMCP, Salesforce Data Cloud, Tealium and CarsXE.

Why care: The best setups follow the same pattern: small scopes, reads before writes, and a person approving anything that changes data. That's the bar.

🔌 New: connect your AI to the forum

Vendors, your AI can now look after your forum thread for you. The forum has its own MCP endpoint:
https://forum.automotivemcp.ai/mcp

Here's how to connect:

Create a forum account at forum.automotivemcp.ai. If you're not signed in, the endpoint sends you to the forum login.
Connect your AI. In Claude or ChatGPT, add the URL as a custom connector and sign in with your forum account. In Claude Code or Cursor, open forum.automotivemcp.ai/mcp/token, approve, and copy your personal token. It's shown once and works for a year, and the page gives you the exact setup command.
Put it to work. Your AI can find your thread, list unanswered questions, read and reply, and post product updates, which get their own "product-update" tag. Try: "Find our thread on the AutomotiveMCP forum and show me any questions nobody has answered yet."

Posts go out under your own name with your normal forum permissions. The endpoint holds no admin key, and you can revoke access anytime from your forum profile. We're connecting our own forum accounts to Claude first, so tell us what breaks.

Why care: Dealers ask how your MCP works in your thread. If your AI can answer there, those questions don't sit unanswered.

📚 The directory

The directory lists 79 dealer tools, and 17 of them have an MCP server. Every listing shows its source and when it was last checked. Don't see your tool? Write to builders@automotivemcp.ai.

🛠️ Built by builders

A single-point store built its own operator agent. John Keohane, VP of operations and partner at Temecula Valley Buick GMC in California, runs an agent he calls "Emily" on Podium's Operator. It follows every customer conversation for him. Every week it runs a "revival campaign": it looks back 45 days of opportunities and sales, removes duplicates, checks which trims he has too many of, and re-contacts unclosed leads with a new incentive or a better trade number. He went from generic bulk texts to messages based on each customer's profile. Emily found the cause of a three-month customer-pay slide in six seconds (a fleet account had moved to mobile service), and he credits it with customer-pay ROs up 12 to 13% this year. His view: BDC work is "low-hanging fruit for what technology can do." His challenge to other dealers: pick one sore spot and let an agent take it. "I think you'll be surprised with how easy it is." (Podium-sponsored segment.) CDG News · Podium transcript

Why care: The win came from leads he already had in his CRM, not new ones he bought. Start there.

A 14-store group now asks every vendor about MCP. Dustin Schuler runs marketing for Bud Clary Auto Group, 14 stores in Washington State. On What's Working in Auto with Jake Hughes, he walks through how his agents rebuild monthly specials and banners for every store. They start at 6am when incentives drop and finish in about 30 minutes, down from "a full day, a day and a half." He still won't build his own CRM, because he wants the SLA, not the tech debt. And when he picks vendors now, the first question is MCP: "it was a nice to have and then now it's like definitely it's becoming more of a requirement" (29:05). Episode · Widewail writeup

Why care: Vendors, if a dealer's AI can't work inside your product, you're about to lose deals to someone whose product it can.

🔌 APIs and MCPs shipped

⭐ Shopify: every store is now agent-ready

Shopify's Fall '26 Edition (Oct 28) opens its catalog, cart and checkout to any AI agent. The parts that matter for dealers:

One MCP for shopping. Shop MCP at https://mcp.shop.com/ lets any agent search 1B+ products, build a cart and check out with Shop Pay. Search needs no API key. Carts and checkout work without an account.
The shopper approves the payment. With Shop sign-in, the agent can use saved addresses and Shop Pay, and it completes the purchase only after the shopper confirms the total and payment method. Shopify says the card is never shared with the agent.
The merchant keeps the customer. Orders land in the merchant's Shopify admin like any other sale, with the customer's name, email and shipping details.
WebMCP on every store. Every Liquid and Hydrogen storefront and checkout now registers tools browser agents can call: search, cart and checkout, with checkout completing only after the buyer approves. No merchant setup is needed.
Merchants approve what agents are told. A new "knowledge base" (Sales channels > Agentic) lets merchants review and publish the facts agents get about their store, like return windows and shipping. Nothing Shopify drafts reaches an agent without their sign-off.

Builder notes: use the root URL (adding /mcp to the path returns a 404), MCP protocol version 2026-07-28, and the skill at https://mcp.shop.com/skill.md. There are no webhooks at launch, so poll get_checkout for status.

Developer news · Merchant news · Shop MCP changelog · WebMCP tools

Why care: This is what "agent-ready by default" looks like when a platform decides it for millions of merchants at once. A car deal is harder (trade, payment, F&I, consent), but buyers will now expect the same: search, a real price, a clear approval step, and the store keeping the customer. Shopify's knowledge base is a good model for dealers too: you approve the facts agents repeat about you.

Google opens Gemini and Search to any MCP server. On Oct 28, Google launched a preview of its Connected Apps Platform. You build a cloud-hosted MCP server, then register, test and submit it in the Google Connected Console, and choose Search, the Gemini app or both. Once it's live you see traffic, latency and error rates. Google's guide says to mark every tool readOnlyHint or destructiveHint. It's a waitlist preview for US developers, with more than 50 connected apps already live. Google · Dev guide

Why care: Annotate your tools. The MCP server a vendor builds for ChatGPT and Claude can now also be submitted to Google.

Agents can now shop Whop businesses through MoonPay's PayBox. Announced Oct 29: agents connected to PayBox (it works in Claude, ChatGPT and Grok) can search Whop stores, prepare an order and pay. The buyer picks the rule: "Always Ask" needs a passkey for every purchase, and "Autonomous" lets the agent spend within limits set in advance, revocable anytime. Whop sellers don't change anything. MoonPay notes that PayBox is run by an unregulated group entity. MoonPay

Why care: "Ask every time" versus "spend up to a limit" is the choice a buyer's agent will bring to a deposit or a service payment.

Microsoft puts a managed agent harness in Foundry. Public preview since Oct 28: Microsoft runs the agent loop and sandbox, and you bring skills and tools, including MCP tools with explicitly configured sign-in and approvals. Harness runtime is free until Dec 1, then billed by usage (model calls are billed as usual). Microsoft

Why care: Groups on Microsoft 365 and Azure get a supported way to run an internal agent against dealer MCPs without building the plumbing.

Mistral releases the weights for Large 4. As of Oct 29, Mistral's trillion-parameter model is downloadable from Hugging Face. Mistral says it ranks first among open models on Artificial Analysis' Coding Agent Index and is pitching it at cyber defense. TechCrunch · Hugging Face

Why care: Big groups that don't want customer data leaving their own servers have one more serious self-hosted option.

A vehicle-history MCP where the human pays. Listed in the MCP registry on Oct 29, Vehicle History by VIN exposes three tools: quote a $9.99 report, create a Stripe checkout link, and read the report once it's paid. Its rules for agents are clear: get an explicit yes before checkout, never touch card data, and summarize only what the report says. MCP: https://www.vehiclehistorybyvin.com/mcp

Why care: It's a clean template for any paid dealer action, like a deposit or a history report: the agent quotes, the person pays.

⭐ Repos worth a look

Shopify/shop (MIT): the shop CLI and agent skill behind Shop MCP, which covers search, cart, checkout and order tracking. A full purchase takes five commands. Why care: It's the most complete public example of agent checkout with a human approval step. Read it before you design yours.
modelcontextprotocol/inspector: the tool Google tells developers to test with before submitting a connected app. Run "npx @modelcontextprotocol/inspector" and call your tools one by one. Why care: It's a five-minute check that your dealer MCP's tools list, run and carry annotations.

⚡ This week

Approve what agents say about your store. Shopify now lets merchants sign off on the facts agents are told. Dealers have no equivalent yet, so do it by hand: check your hours, doc fee, return and price policies, and service menu on your site and Google Business Profile. Action: Pick one person to own "what AI says about us" and have them ask ChatGPT, Gemini and Claude about your store this week.
Ask your vendors whether their MCP tools are annotated. At MCPCon (Oct 22 and 23), OpenAI said it rejects ChatGPT plugin servers unless every tool is annotated, and now Google's guide asks for it too. Action: Ask each vendor with an MCP server two questions: which tools are read-only, and which ones change data or contact a customer?
Decide your approval rules for payments. Shopify, MoonPay and Vehicle History by VIN all keep a person approving the payment. Action: Write down what an agent may do without a person (look up inventory, quote, hold an appointment) and what always needs the customer's or a manager's approval (deposits, credit pulls, price changes).
Reynolds goes all-in on one unified agent. President and acting CEO Chris Walsh described a single AI agent, Rey, running across every department on a cleaned-up, deduplicated data layer. Autosphere Action: If you're on Reynolds, ask whether outside agents and your other vendors can reach Rey's data and actions through MCP or a documented API, and on what terms.
Mine the leads you already have. Temecula's win came from leads already sitting in the CRM (see Built by builders). Action: Pull your last 45 days of unsold opportunities and run one personalized re-contact this week, by agent or by hand.

📐 The open standard

Shopify could make millions of stores agent-ready at once because it owns the platform. A dealership doesn't have one platform. Inventory, price, trade, payment, service and consent sit in different vendors' tools. That's why AutomotiveMCP is an open, industry-wide effort and not one company's product: one shared front door, versioned schemas, tools labeled read-only or write, and a handshake so the customer's consent travels with them.

We're also glad to see others pushing the same way. Open Auto went public Oct 20 as a neutral group working on shared standards for interoperability, identity, consent and permissions, and says it's working with STAR on how existing standards evolve for AI (coverage). More open work means less lock-in for dealers, and we'll compare notes.

Spec v0.1 is an open RFC, with versioned schemas for inventory, leads, service and deals. Read it and push back at automotivemcp.ai/spec. Vendors: add your tool to the test kit by URL. Builders: post what worked and what broke on the forum.

Talk soon,
Alex Oleynik, David Ortiz, Brian Pasch, Gray Scott and Daniel Kim
The AutomotiveMCP team

Want the Brief in your inbox? Sign up at automotivemcp.ai.