AI Agent Tools Stopped Being a Proposal in August. One Thing We Told Dealers in July Is Now Wrong.

In the first four weeks of August, three companies turned WebMCP from a proposal into working plumbing.

On August 5, Shopify shipped WebMCP tools live on every Liquid storefront and on the Hydrogen developer preview. Nothing to install, nothing to configure, no merchant decision required. The tools let a shopper's AI agent search the catalog, manage the cart, and go to checkout on the shopper's behalf, through named functions like search_catalog, update_cart and proceed_to_checkout.

On August 6, Cloudflare shipped a WebMCP developer preview that injects the plumbing at the network edge. Their words: "Switch it on and browser agents can start working with your site, with no code and nothing changed at your origin." It is a toggle in the Cloudflare dashboard under Agent Readiness, and the next HTML your site sends carries the bridge.

On August 25, OpenAI turned it on inside ChatGPT. "We're adding support for WebMCP in the ChatGPT desktop app's built-in browser and ChatGPT Sites," the company posted. "When you visit a compatible website, ChatGPT or Codex can automatically use it to complete your task."

If you have never heard of WebMCP, the one-line version is this. It lets a website hand the browser a menu of named actions, each with a description and a structured input schema, so an AI agent calls a function instead of squinting at your page and guessing which button to click. OpenAI's framing in its own documentation is blunter: "Instead of leaving agents to guess their way through your UI, you define exactly how they can use your app."

We wrote about this in July, and one thing we told you is now wrong

On July 10 we published a piece arguing that WebMCP changes the question from "can AI read my site" to "can AI use it". The thesis held up. One of the action items did not.

We told dealers this: "Most dealer sites run on a vendor platform, so you can't add these tools yourself, but you can find out whether the company you pay is watching this or asleep on it."

Cloudflare's August 6 release is a direct counterexample to the first half of that sentence. If your dealership site sits behind Cloudflare, the bridge goes on from a dashboard toggle. Cloudflare uses HTMLRewriter to add one line to each HTML response, a reference to a script at /.webmcp/bridge.js. Your website vendor never touches their codebase. That is a genuinely different situation from the one we described seven weeks ago, and dealers who took our July advice as "wait for your vendor" deserve the update.

We also framed the whole thing in July as an experiment with nothing shipped to end users. Three weeks later OpenAI is executing site tools inside ChatGPT's own browser. The proposal-versus-product line moved.

Why this pattern should look familiar

We have watched this exact sequence three times now. A new kind of visitor shows up on the web. The platforms define what "ready" means for that visitor. A window opens where being early is cheap and being late is expensive. Then readiness becomes table stakes and nobody gets credit for it.

Mobile-friendly ran that course a decade ago. Structured data ran it after Schema.org. Both times, the dealers who moved during the window spent very little, and the ones who waited spent years and real money catching up while their competitors sat on the rankings.

What is different this time is the speed and who is setting the pace. Retail moved first and moved by default. Shopify did not ask merchants to opt in. Millions of storefronts woke up on August 5 with a working machine-readable cash register. Expedia, Instacart and Target are in the same conversation. Automotive retail is not in that list.

What this means for your dealership

Picture the same shopper we described in July. She tells her assistant to find a certified CR-V under $30,000 within forty miles, get the out-the-door price, and book a test drive Saturday morning.

On a Shopify store, the agent has a defined path for the retail half of that job. On your dealership site, it has to improvise: parse an inventory grid that renders through JavaScript, work around a "call for price" gate, and fill a test drive form built for a human's fingers. When the agent stalls, you get no notification and no lost-lead report. The shopper just sees a different answer.

Here is the sharper version of the problem. The gap that opened in August is a gap in reliability, and reliability is what an agent optimizes for. An agent that can complete the task on one site and only sometimes complete it on another will route toward the first one. Nobody at Google or OpenAI has to decide to penalize your store. The agent's own success rate does it.

Now the part that keeps this honest, because there is a vendor pitch coming and you should be able to see through it.

No dealership is losing a sale to WebMCP today. The specification is still a Draft Community Group Report from the W3C Web Machine Learning Community Group, dated August 26, and it says plainly that it "is not a W3C Standard nor is it on the W3C Standards Track." Shopify notes that agent support is currently limited to Chromium-based browsers. Google's implementation is an origin trial from Chrome 149. OpenAI's version reaches ChatGPT Work and Codex users on GPT-5.6 Sol or Terra, with Luna excluded and Enterprise and Edu workspaces excluded too. That is a developer and business audience, well short of the mass consumer tier.

The Cloudflare toggle deserves the same scrutiny. The developer preview ships two tool packs, Content Credentials and Site MCP Server. Neither one gives an agent a book_test_drive function or a real out-the-door price. Cloudflare is candid about the limits of its own preview: the Content Credentials pack reads and reports a credential rather than cryptographically verifying it, and returns signatureVerified: false so that agents avoid "mistaking a decoded claim for a checked one." Flipping that switch makes your site addressable. It does not make your site useful.

And there is a security dimension that any dealer principal should hear before a vendor sells them "agent readiness." OpenAI's own documentation warns that "Website-provided tool definitions and results are untrusted content. A tool's name or claim that it only reads data isn't proof of what it does." OpenAI runs a safety review on each tool invocation and still says that "these checks reduce risk; they don't make a website or its output trustworthy." The platforms are treating exposed site tools as a live attack surface. Anyone who sells you WebMCP without mentioning that has not read the documentation.

What to actually do about it

1. Find out whether your site sits behind Cloudflare, and who holds that account. This is a five minute question for your IT or website contact. If the answer is your website vendor, you have learned something important about how much of your own infrastructure you control.
2. Do not flip the Cloudflare toggle on your production site this week. It is a developer preview with two generic tool packs. Know that the option exists and what it costs, then decide deliberately rather than because a rep asked.
3. Ask your website vendor a sharper question than we gave you in July. "What is your plan for WebMCP" invites a brochure answer. Ask instead: "Which specific actions on our site will you expose as tools, and on what date?" Inventory search, pricing, trade appraisal, credit pre-qualification, and service scheduling are the five that matter. A vendor who cannot name a date is telling you the honest answer.
4. Fix the underlying actions first, because a tool only exposes something that already works. Go use your own site the way an agent would. Can you reach a real price without a form gate? Can you book a service appointment end to end? Can you start a trade appraisal without hitting a dead field? If you cannot finish, exposing that flow as a tool just publishes the failure in a machine-readable format.
5. Get your inventory data clean regardless. Structured vehicle records with year, trim, mileage, drivetrain and an actual price are the input to every version of this, the citation version and the agent version alike. This is the work that pays off no matter which standard wins.
6. Watch the calendar, not the hype. OpenAI is running a WebMCP Challenge with submissions closing September 3 and winners announced September 23, backed by Shopify, Google Chrome, Cloudflare, Netlify, Vercel and Render. That tells you which way six large platforms are leaning. Watch what those companies ship in the fall.

The part worth remembering

For most of the last two years, the AI conversation in this industry has been about visibility. Getting cited. Getting found. Getting named in the answer. That work is real and we have written about it constantly.

August added a second scoreboard. Being found puts your store in the consideration set. Being able to complete the shopper's task is what turns that into an appointment, and starting this month a meaningful slice of the web can do the second part while dealership websites still cannot.

The window is open and it is cheap right now. Most of what it asks for is the same homework we have been pointing at for two years: real prices, clean inventory data, and forms that actually work. If you want a second set of eyes on whether your site could survive an agent trying to book a test drive on it, book a walkthrough and we will run it with you.