In March 2026, OpenAI pulled the plug on Instant Checkout. Barely 30 Shopify merchants were actually connected, no inventory handling, no sales tax. Dead on arrival. And everybody read that as proof that buying through AI agents is vapor.
Wrong signal.
TLDR: Agent checkout does not work yet. Everyone reads that as failure, and it is the wrong signal: agent-driven discovery already jumped 15x. I built 2 MCP servers this year, and that road showed me what is waiting for anyone who wants to sell to agents.
Checkout breaking on the first payments, that is not a contradiction. It is the signature of a tech in its very first days. Part of it already works hard. The rest is held together with tape.
We have seen this movie before. In 1996, the web was half broken too, and it changed everything anyway.
In 1996, Your Website Didn't Work Either
In 1996, the web did not work. I mean that literally. Pages loaded a row of pixels at a time, and the connection dropped if someone else in the house picked up the phone. Half the sites were hand-built in HTML that would make you cry today, the other half came out of Dreamweaver looking like a ransom note. Paying for something online felt reckless, like handing your wallet to a stranger because he promised to mail it back.
It was bad. And it was already not optional.
The companies that looked at that mess and said "call me when it actually works" were not wrong about the mess. They were wrong about what the mess meant. So they waited. They told themselves the customers were not online yet, the tech was not ready, the real version was a few years out. The proof showed up around 2005, dressed as everyone else's e-commerce store and eating their lunch, and by then they were 10 years behind and paying agencies a fortune just to catch up to the baseline.
We are standing in that exact spot right now. Swap websites for agentic commerce and you get the same picture: it looks ugly, it half works, and smart people who should know better are saying it can't be serious.
Everyone's Reading the Wrong Signal
There are 2 camps arguing about agentic commerce, and both of them are looking at the wrong gauge.
The skeptics point at Instant Checkout and call the time of death. They are not wrong about the facts. OpenAI shut it down in March 2026 with around 30 Shopify merchants actually live, no real inventory handling, no sales tax logic. eBay banned third-party buy-for-me agents outright. Surveys put maybe 17% of shoppers comfortable letting an AI actually complete a purchase. If that is the whole picture, sure, go ahead and bury it.
The hype camp does the mirror version. They wave analyst decks with trillion-dollar numbers printed on them and act like the future already shipped.
Both of them are staring at checkout. The signal is not at checkout. It is at discovery, and discovery is not a projection sitting in a slide, it is already on the dashboard. Shopify says orders coming in from AI search jumped 15x since January 2025. Adobe clocked generative-AI traffic to US retail up 4700% year over year. That is not a forecast you have to buy into. That is shoppers already starting their purchase inside an AI, then finishing it somewhere. The front of the funnel moved. It moved last year.
The part both camps skip is simple. A thing breaking is not the same as a thing being a bad idea. Instant Checkout did not fail because agents buying things is a dumb concept. It failed because the plumbing under it was never built. Forrester put it plainly, the bottleneck moved off payments and onto product data. The Drum, looking at the same shutdown, called it an execution failure rather than a conceptual one. Even Fast Company, talking to the people actually building this at Google and OpenAI, came back with months, not years. So "it broke" is not the opposite of "this is the moment." It is the same observation said twice.
Now the honesty tax, because this topic earns it. I run 2 MCP servers right now and both of them are internal tooling, the merchant-side kind. They let me drive my own back office from my own agent. Neither one is a storefront that a stranger's shopping agent can walk into and buy from. So I am not coming at this from inside the agentic-commerce gold rush. I am coming at it from the build path that leads up to it, and the only reason that is worth your time is that the build path already showed me in concrete detail what is going to land on anyone who does want to sell to agents. That is the entire reason this article exists, and also why I am not going to pretend I have a working agent storefront when I plainly do not.
An Agent Reads Your Structure, Not Your Marketing
Strip everything else away for a second. When an AI agent lands on your business, what does it actually perceive?
It does not see your landing page. It does not see the headline you A/B tested 40 times, or the brand colors, or the testimonial carousel, or the founder photo on the about page. An agent does not read marketing. It reads structure. It hits an endpoint and either gets back clean machine-readable data or it gets back nothing it can use.
That is the whole mechanism, and it reframes the entire problem. Getting ready for agentic commerce is a plumbing problem, not a marketing one. The work is exposing your data in a shape a machine can read. Your beautiful product page can stay beautiful for the humans. The agent needs the same facts (the price, the variant, the stock count, the shipping rule) sitting behind an endpoint as structured data. If that endpoint does not exist, you are simply not in the room when the agent makes its shortlist. Everything else here, the build story and the two paths, comes straight out of that.
There are 2 flavors of this and they are easy to mix up. Consumer-side MCP is when somebody else's shopping agent calls your catalog from the outside, scoped to your store. Merchant-side MCP is when you, the operator, drive your own business from your own agent. I have built the merchant-side kind. The consumer-side kind is the one with the gold rush attached, and it is still mostly empty real estate. Worth knowing too: MCP is the discovery and tool-calling layer, the part that lets an agent find you and read you. The actual payment handshake rides on other protocols stacked on top, and that layer is genuinely still being fought over.
Now, regular readers are about to catch me in what looks like a contradiction. My most-read piece argues that for wiring up your own agent's tooling, you should skip MCP and reach for plain CLIs instead. And here I am telling you MCP is the thing. Both are true, because they answer different questions. Your agent doing work for you is a different problem from the outside world's agents finding and reading you. If you want the long version of why the internal job goes the other way, I laid out the full case for CLIs over MCP. For the storefront job, the boring structured version is the one that wins.
What My Own MCP Server Actually Cost
Time to make "held together with tape" concrete, because vague hand-waving is the hype camp's whole move, and I would rather just show you the tape.
February 25 this year. I am putting my own MCP server online. Stack is Next.js and Convex, deployed on Vercel. On paper, a 1-hour job.
It was not a 1-hour job.
First, serverless. Vercel functions are stateless and short-lived, so anything I had written that assumed a session sticking around had to be ripped out and rebuilt stateless. Then the dynamic route just broke on Vercel. First request fine, every request after it a clean 404, which is a fun one to debug because it works perfectly on your own machine. Then the connector itself. I had a custom Bearer token auth that worked great with curl and that the Claude.ai connector flatly refused to recognize.
So I got to build OAuth 2.1. From scratch. Self-hosted, signed JWTs (JSON Web Tokens), no database, because adding a database just to hold sessions for a single-user tool is its own special kind of madness. Halfway through, jose, the JWT library, crashed Vercel's bundler. Just refused to build. (Vercel's logs will not point at jose either, you find that out by deleting things until it works again.) That was a useful little humiliation, the kind that reminds you the whole ecosystem is younger than the blog posts make it sound.
And the finish line: my access tokens kept expiring, Claude.ai dropped the connection every 24 hours, so I set the token expiry to 100 years. It is ugly. It is also fine, because this thing has exactly 1 user and his name is me. (The proper fix is a refresh-token dance. Yes I know. No I have not done it. It has not bitten me in 3 months.)
1 honest caveat before you panic. That entire mess is my stack. Next.js, Convex, Vercel, Claude.ai as the client. Pick a different stack, or connect to a hosted MCP server that somebody else already operates, and you hit a completely different set of walls, or almost none at all. My war story is a situated experience, not the one true path. Read it as the shape of real friction, not as proof that this is too hard. Because there are much shorter paths.
None of that was science fiction. It was friction. Annoying, dated friction, the kind where the Stack Overflow answer does not exist yet. It is 2026 and this stuff is still held together with tape.
That is not the reason to wait.
The tape is the tell.
Where to Start Right Now
So where does someone actually start. There are 2 doors here, and the title said wire up, not build from scratch, on purpose. Both doors are legitimate.
The fast door: borrow the infrastructure
If you do not want to hand-write an auth layer, good news, you mostly do not have to. The money layer already shipped. Stripe runs a hosted MCP server at mcp.stripe.com, and connecting to it is an OAuth click that takes seconds. Every Shopify store already exposes an /api/mcp endpoint for public catalog operations, no auth needed on the read side. PayPal shipped its own. WooCommerce has one in public beta. The pattern is the same everywhere: somebody else already operates the server, you connect to it, and your data becomes machine-readable to an agent in minutes.
For most people reading this, that is the entire move. You are not behind because you have not built anything. You are behind if your catalog, your prices, and your availability are sitting in a format that only a human with working eyes can parse.
I connected the Stripe server for the first time sitting by the pool with the kids doing cannonballs 2 meters away, and the thing just worked. OAuth click, done. I remember being almost annoyed at how anticlimactic it felt after the month I had sunk into my own server. The kids did not care about any of that. They wanted me to rate the cannonballs. So I rated the cannonballs.
The control door: run your own
The other door is the one I went through, and you already read what it costs. You build and operate the server yourself. The upside is real: total control, plus the ability to expose logic that no generic catalog endpoint covers, your weird pricing rules, the bundle math, those partner-specific feeds. It is not a scarecrow I am putting up to push you toward the easy door. It is a legitimate choice for anyone whose business does not fit the shape that Stripe and Shopify pre-cut for you. It just costs you the war story.
And the cost nobody prints on the label: agentic latency. People call it that, and it is exactly what it sounds like. If your hosted server is slow, the agent calling it is slow, and a slow agent is one that gets dropped halfway through the task. The fast door trades control for somebody else's uptime. The control door trades your evenings for the ability to make it fast on purpose. Pick the trade you can actually live with.
For the non-dev crowd this piece is aimed at, start with the fast door. Connect the existing infrastructure, then go and check whether agents are actually reading your data, whether anything is hitting those endpoints at all. Internalize later, if and when the need is real. And I will hedge that, because I am genuinely not sure the run-your-own path scales past a solo operator without quietly turning into a part-time job. Solo, it is fine. For a team, I would want to watch it run for a while before I promised you it holds.
If you want that whole path written out step by step, I put it in my book on going from demo to live app. And if the control door is calling you but the war story spooked you, the gentler on-ramp is to self-host something small first. I wrote up putting my own agent on a $5 server, and it is a far softer introduction to operating your own thing than jumping straight into OAuth 2.1.
We Are in 1996 Again
I still remember my 14k modem. The internet back then was a slog, and you could actually hear it negotiating. And I remember the small jolt of fear the first time I bought something online, my card number going into a box I had no real reason to trust. (My mother-in-law still phones us when she wants to buy something on the web. She will not put her card in herself. In 2026.)
Same thing now.
You will feel that same small jolt the first time your agent buys something for you. (And if you asked it for Tokyo and it books you Ouagadougou, well, now you know there are still a few bugs to iron out 😬.)
Tomorrow, every business will have an agent or an MCP connector the way it has a phone number and a website today. McKinsey is throwing around $1 trillion of agentic commerce by 2030. eMarketer already counts $20.9 billion in US retail spending through AI platforms for 2026, roughly 4x last year.
The window to position yourself is right now. Not when it is industrialized. Not when it is clean.
Waiting for proof is just missing 1996 a second time.
Sources
- OpenAI Just Killed Native Checkout in ChatGPT: Paz.ai, the Instant Checkout shutdown and the Forrester read on product data as the new bottleneck
- Don't Dance on the Grave of Agentic Shopping Just Yet: The Drum, execution failure versus conceptual failure
- Shop 'til You Bot: Fast Company, the "months not years" reporting from inside Google and OpenAI
- Agentic Commerce Playbook 2026: EcomExpo, the Shopify 15x figure and the eMarketer spending numbers
- MCP Servers for Ecommerce: The 2026 Developer's Guide: YNS, consumer-side versus merchant-side MCP
- Best MCP Servers for Ecommerce: Checkout Page, the hosted server roundup and agentic latency
Top comments (0)