A customer came through my Messenger support chat a couple of weeks ago. They run an Australian e-commerce site, selling children’s insoles as a subscription, and they wanted to add a feature that doesn’t exist as a product in WooCommerce today. They needed a checkout dropdown where the customer picks a registered school or club to receive a five-dollar donation on every order. They described themselves as a non-developer with an ex-developer background, time poor, budget conscious. They were not going to cobble this together out of ten plugins and a Zapier flow. There was no plugin that did it, because what they wanted was genuinely novel.

I told them honestly this was going to need a small extension for Siren to do, pointed them at Siren’s free tier and a Claude subscription, and told them to connect Beacon (Siren’s MCP server) to Claude Web. They came back the next day with a working plugin. They successfully shipped a real Siren extension, the kind of thing a consultant would have charged them four figures to build, in an evening. Under an hour to build, thirty to forty-five minutes to wire it up and test. A non-developer.

The thing I keep turning over is that the MCP server wasn’t just the tool that helped them build it. It was also part of the reason they trusted Siren before they trusted me. By the time we talked, Claude had already walked them through what Siren was, what parts of the problem Siren could handle natively, and what parts would need extension. They didn’t need me to sell them on it. They needed me to confirm the path Beacon had shown them was the right one.

I didn’t see that coming when I shipped Beacon. Product discovery for Siren is moving into AI assistants. Not metaphorically, not soon. It already has.

How Siren’s MCP Server Sells Siren

I built Beacon because the customers who come to Siren mostly aren’t coming with cookie-cutter requests. They’re coming with novel concepts: “I want to pay my course instructors a royalty on every enrollment, and run an affiliate program on the same catalog, and the two shouldn’t double-pay on the same sale.”

Or “I want customers to pick which nonprofit gets a donation at checkout and pay the nonprofit monthly.”

Or “I want a tiered sales bonus where the top performer each month gets a larger cut, and the runner-up gets a smaller cut, and everyone below gets a flat rate.”

Siren is genuinely capable of handling all of those, but you have to know which pieces to combine. That’s a lot to ask of a prospect who’s evaluating you on a sales call.

Hot take: I hate onboarding flows with a heated passion. They interrupt you when you know what you’re doing, and they’re only helpful the first time. I wasn’t going to build another wizard. I wanted prospects to complete the thinking part of setup by talking to an adviser who could explain things in plain language and point at real configuration examples. MCP was the obvious shape for that, because Claude, ChatGPT, and Cursor were already where my buyers were spending their time.

The case I presented earlier is the cleanest proof I have that it works. The customer came in with a novel concept. Beacon walked them through what Siren could and couldn’t do natively. Beacon pointed Claude at the engagement triggers documentation once their first attempt tried to route around Siren’s extension system. Their second attempt, generated from that same Claude session, was architecturally correct. They dropped it onto their site after testing it in a safe environment, and it worked.

That whole loop happened without me scheduling a call, writing a proposal, or pitching anything. By the time we talked, the sale was effectively already made. They weren’t asking whether Siren could solve their problem. They were asking me to confirm the path Beacon had already mapped out was going to do what they wanted it to do.

MCP is today’s version of a bigger shift

Right now, in this millisecond of time in the crazy constant shifting of AI, MCP is a key piece of the technology that makes this accessible to less-technical users. I’m sure in-time this will shift into more native systems, however.

Cloudflare released isitagentready.com earlier this month, which scores how well your website exposes itself to AI agents natively (no MCP required). Siren’s site scores a hundred there right now, and that score matters. Some of the questions agents ask about Siren will eventually be answerable from the site alone, with no MCP connection needed. That’s the direction the whole ecosystem is pointed.

Today, though, there’s no contest. I’ve tested both paths with the actual questions my customers ask, and the MCP server consistently performs much better than the web-native path for anything involving program design, recipe generation, or extension guidance. Agent-friendly websites are a floor, not a ceiling. MCP is where the elevated experience lives today.

The real claim isn’t “build an MCP server, specifically.” The real claim is: the conversation where a buyer decides whether your product fits them is increasingly happening inside an AI tool, and you need to be present in that conversation. MCP is how that presence is delivered today. If the mechanism changes in two years, the observation still stands.

Websites are a given, and AI-optimized tooling like MCP servers are next

Rewind to around 2005. If you ran a business without a website, you were unusual. Not dead, but unusual. You were accepting that a chunk of your discoverable market couldn’t find you unless they already knew your name. By 2010, no website meant invisible to a generation of buyers trained to Google everything.

AI-optimized tooling is in that same early window right now. Not having any doesn’t kill you today. You still get found through Google, word of mouth, content, ads. But a growing percentage of your buyers are starting their evaluation inside an AI assistant, and that percentage goes up every month. In two or three years, “I don’t have anything an agent can talk to” will sound the way “I don’t have a website” sounded in 2010.

The founders who are going to do well here are the ones who stop treating agent-facing surfaces as a roadmap item and start treating them as the front door for a channel they don’t yet understand. Not because the channel is hypothetical. It’s live. My buyers are already in it. The customer I described earlier was in it before they ever messaged me. If yours aren’t, it’s probably because they’re in your competitors’ MCP instead, and you don’t know it, because nobody’s going to tell you “I decided against you based on a Claude conversation.”

Build the thing. Keep it small. Treat it like a surface your customers will actually use, not a REST mirror. And accept that the first impression of your product, for a real and growing share of your buyers, is going to be made without you in the room. Like it was made without me in the room when that customer spent an evening designing their donation-at-checkout flow with Claude.

That used to be what the landing page was for, but that conversation has begun to shift off your website. AI-native tooling like MCP servers make it possible for that to happen.