Videos

We Built a Store That AI Agents Can Shop: Inside Wardenclyffe Engine

8:22 · November 20, 2026 · Watch on YouTube

A walk through Wardenclyffe Engine, the commerce engine ProjectThunder built for people and for AI agents. One set of business rules sits behind a server-rendered storefront, a control room with 99 admin screens, a REST API and 50 MCP tools for agents. The video covers agent checkout in four calls, the technical specifications measured in October 2026, the ERPNext and QuickBooks Online add-ons, how two people built it with AI and 1,240 tests, and the Business Source License. The ACP and UCP checkout adapters are built and tested but not yet live.

Read the full write-up: Wardenclyffe Engine

Chapters

  • 0:00Cold open
  • 0:23What it is and why we built it
  • 1:14The storefront
  • 1:52The control room
  • 2:45One core, three doors
  • 3:21Built for AI agents
  • 4:43Technical specifications
  • 5:43Add-ons and integrations
  • 6:50How it was built
  • 7:40Licensing and close

Shorts from this video

Transcript

Cold open

A store used to have one kind of customer: a person with a browser. That is changing. More and more, the first visitor is an AI agent, sent by a real buyer to find a product, compare it, and place the order. Most commerce software was never built for that visitor. So we built one that is. This is Wardenclyffe Engine.

What it is and why we built it

Wardenclyffe is a complete commerce engine. It has a storefront for shoppers, a control room for the people who run the store, an API for other systems, and a tool interface for AI agents. All four sit on one set of business rules. The name comes from Tesla's tower on Long Island, a station built to broadcast. That is the idea here: a store that sends a clear signal to anything that wants to do business with it. We have been building commerce software since 2004. Our own system had twenty years of rules in it, and every one of them assumed a human on the other end. So in May of 2026 we started again. We read all of the old code, about a hundred and thirteen thousand lines, as a specification. Then we built a new engine around what it taught us.

The storefront

Start where the shopper starts. The storefront is rendered on the server, so the price in the page is the price. That matters when the visitor reading the page is not a browser. Shoppers get faceted browsing with live counts, search, sale pricing, reviews, and collections, which are editorial pages with a hero, a story, and hand-picked products. Each one has its own address, its own entry in the sitemap, and its own share tags. The look is a theme. What you are seeing is our demo theme, an atelier corsetry brand. The same engine can wear a completely different brand without touching the code underneath.

The control room

Now the other side of the counter. We call the admin the control room, and it is styled after the tower: brass, copper, and night blue. There are ninety-nine admin screens, grouped into catalog, operations, marketing, and system. Live gauges sit at the top. Every operating surface is one click below them. The catalog handles products, variants, brands, categories, suppliers, custom fields, and fitment data, for stores where a part has to match a model. Pricing has customer groups and volume tiers. And there is a bulk import built for the systems that already feed your catalog. It takes compressed files up to two hundred megabytes, loads them in batches of ten thousand rows, and is safe to run twice. Every change an administrator makes is written to an audit log: who did it, what changed, and what it looked like before.

One core, three doors

Here is the design decision that shapes everything else. An AI agent is a client. It is not a bolt-on integration. Every capability in the engine is written once. Then it is offered through three doors: a page for a person, a JSON endpoint for another system, and a tool call for an agent. All three run the same code. That removes the most expensive kind of bug in business software: the one where the web form and the API disagree about the rules, and nobody notices for months. Here there is no second place for the rules to live.

Built for AI agents

So what can an agent actually do? The engine speaks the Model Context Protocol, MCP, which is the standard way AI assistants call tools. There are two routes. One is for a customer's agent, which can search, build a cart, and work with that customer's own orders. The other is for an operator's agent, which can manage the store. Between them there are fifty tools, and each tool checks the caller's own permissions. An agent can do what its owner is allowed to do, and nothing more. The engine also publishes its own tool registry, so an agent can learn what the store offers without anyone reading the source. AI shopping does not have one standard yet. So beyond MCP, we have built adapters for two checkout protocols: the Agentic Commerce Protocol, and the Universal Commerce Protocol. In our test runs, an agent placed an order in four calls. It created a checkout, added the buyer and shipping details, and completed the order. When the same request was sent twice by mistake, the engine replayed its first answer instead of placing a second order. The totals matched to the cent. One thing to be clear about. Those two checkout adapters are built and tested, and they are switched off by default. They are not yet running in any public AI shopping surface. When that changes, we will show it.

Technical specifications

For the engineers watching, here are the specifications, as measured in October of 2026. The engine is written in C# on .NET, with server-rendered Razor Pages and Entity Framework Core. It runs on SQLite for development and on SQL Server in production, from the same data model. The code is about one hundred and thirty-seven thousand lines, in a layered solution: a domain layer, an application layer, infrastructure, the web application, and a command line tool. It exposes two hundred routes: one hundred and thirty-three pages, and sixty-seven API endpoints. The data model has one hundred and twenty-two entities in three separate databases, for commerce, operations, and identity, with thirty-eight migrations. Search is built in, using Lucene, so there is no separate search service to run. And one command gives a developer a full working store on a laptop.

Add-ons and integrations

A store does not live alone, so the engine comes with add-ons. First, ERP accounting. There is a connector layer that is not tied to one vendor, and a first connector for ERPNext. It syncs products, SKUs, prices, customers, suppliers, and warehouses. Pushing web orders into the ERP is the piece we are finishing now. Second, accounting. There is an adapter for QuickBooks Online that sends customers, products, and orders. Third, payments. The engine has one payment interface, covering authorize, capture, refund, and void, with adapters for major payment providers. Every payment step is recorded, and a retried request never becomes a second charge. There is more. Shipping methods and tax rates by region. Discounts, customer groups, and a loyalty program. Returns. Abandoned cart recovery. Email notifications and a newsletter. An AI assistant with usage and cost tracking. And a command line tool to create, migrate, seed, and check a store.

How it was built

Two people built this in about five months. AI wrote most of the lines. People made every decision. That only works with a strong checking layer. There are one thousand two hundred and forty automated tests, and the full suite runs in under a minute. For every fix, we break the code on purpose, and the new test has to fail. If it stays green, the test is wrong, and we rewrite it. We also learned not to take a review at face value. In one large automated review, twenty reviewers raised one hundred and forty-nine findings. Eight were real. Finding problems is cheap. Proving which ones are true is the work. And every build checks that the data model and its migration scripts still agree, in all three databases. Nobody can change the schema and forget the migration.

Licensing and close

Wardenclyffe is source-available under the Business Source License. It is free to run in production for stores under one million dollars in annual sales, and every release becomes Apache 2.0 in May of 2029. It is self-hosted. Your store, your servers, your data. We built this engine for ourselves, which means we can show you everything: the code, the tests, and the numbers behind this video. If you have a system that needs to take its next step, that is the work we do. We are ProjectThunder. Tell us where you are, and where you want to be. The link is in the description. Thanks for watching.

Narration is an AI voice (Azure). The facts come from the linked write-up.

More videos