Everyone's talking about MCP servers as a developer tool. Connect your IDE to a database. Give your coding agent access to GitHub. Wire up Jira so Claude can read your tickets. That's the story so far, and it's valid; but it's also just chapter one.
The chapter is the one where MCP servers become the way businesses reach customers. Not through a website. Not through an app. Through the AI assistant the customer is already talking to.
We just built one. And I recently used several to plan a vacation that would have taken me days to organize manually. Both experiences taught me the same thing: MCP is quietly becoming the most powerful distribution channel nobody is marketing correctly.
What We Built: Hi Mami MCP
At Enpitech, we recently built an MCP server for Hi Mami (הי מאמי) one of Israel's leading deals, coupons, and benefits platforms with over 115K followers on Instagram. The kind of platform where you go to find out which brand has a 1+1 this week or where to get 30% off your next online order.
The MCP server is technically straightforward: search and fetch. You can search for brands, browse campaigns, look up product categories. What makes it different is what the user sees. Instead of raw JSON that the AI has to interpret and reformat, the server returns rich interactive UI cards, the same visual language as the Hi Mami website itself. Beautiful deal cards with discount badges, brand logos, coupon codes, and direct action links. Theme-aware, RTL Hebrew support, rendered natively inside the AI chat.
This is the part most people miss about MCP. It's not just a data pipe. With MCP UI responses, you can deliver a branded, visual experience directly inside Claude, ChatGPT, or any compatible assistant. The customer never leaves the conversation. They ask "איזה מבצעים יש היום בקסטרו?" - “What sales can we find today on Castro?” and get a cards with deals, the code, and a link to redeem, right there in the chat.
That's a storefront.
What I Used: Planning a Vacation With MCP
Here's the other side of the story, using MCP servers as a consumer.
I recently had to plan a complex vacation where three families from three different countries needed to meet in one place. Each family flying from a different city in Europe, with different budget constraints, different schedule flexibility, and kids involved. This is the kind of trip that usually involves twenty browser tabs, three spreadsheets, and a group chat that goes on for two weeks.
Instead, I connected Claude to flight aggregator MCPs and Google Maps. I gave the agent the constraints: three origin cities, date ranges, budget limits per family, preference for direct flights where possible, maximum acceptable connection time. Then I let it work.
Claude compared flights from multiple airports across different countries. It calculated connection times, evaluated route options, cross-referenced pricing, and found overlapping windows where all three families could arrive within hours of each other. It considered secondary airports that I wouldn't have thought to check. It factored in travel time from each airport to the destination.
The result was a complete travel plan that balanced cost, convenience, and timing across three families in three countries, something that would have taken me the better part of a week to coordinate manually. It took an afternoon.
That experience changed how I think about MCP. This wasn't a developer workflow. This was a consumer solving a real problem by giving an AI agent access to the right tools.
The Bigger Picture: MCP as a Distribution Channel
Here's what both experiences taught me. The Hi Mami MCP puts a brand inside the AI conversation at the moment a customer is looking for deals. The flight MCPs put travel services at my fingertips when I was actively trying to book. In both cases, the MCP server wasn't a technical integration, it was a presence in the space where decisions get made.
This is the shift that businesses need to understand. The primary interface between a customer and a business is changing. Today, millions of people start their day by asking an AI assistant a question. When that question is "where can I find a good deal on running shoes" or "help me plan a trip to Barcelona," the businesses that have MCP servers are the ones the agent can surface. The ones that don't are invisible.
Think about what that means for a deals platform like Hi Mami. Instead of hoping a customer opens the app or visits the website, their deals show up inside any conversation where someone asks about shopping, discounts, or specific brands. The MCP server becomes a distribution channel that meets customers wherever they already are.
And this isn't speculative. The MCP ecosystem grew from a few hundred servers when Anthropic released the spec in November 2024 to over 17,000 indexed servers by early 2026. Every major commerce platform has shipped support. Shopify has four official MCP servers. Stripe runs one at mcp.stripe.com. The infrastructure is here.
Why Most MCP Marketing Misses the Point
Here's the problem: almost every MCP server today is marketed to developers. The GitHub READMEs talk about installation, configuration, and API endpoints. The landing pages say "connect your AI agent to our platform." That's fine for developer tools, but it completely misses the consumer-facing opportunity.
If you're a business with a customer-facing product — deals, travel, food delivery, real estate, anything where people search and transact — your MCP server marketing should answer one question: what can your customer do with this that they couldn't do before?
For Hi Mami, the answer is: "Ask any AI assistant for today's deals and get beautiful, actionable results without opening an app." For a flight aggregator: "Tell your AI where three families need to meet and let it do the work." For a restaurant booking service: "Say 'find me a table for four tonight near Rothschild' and get options you can book from the chat."
The MCP server is the product's ambassador inside the AI conversation. Market it like one.
What Comes Next
I believe we're about to see a wave of consumer-facing MCP servers that changes how businesses think about distribution. A few predictions:
MCP as SEO for the AI era. Just like businesses invested in being discoverable on Google, they'll need to be discoverable inside AI assistants. Having an MCP server will become as essential as having a website.
Rich UI becomes the differentiator. The Hi Mami MCP stands out because it returns visual cards, not raw text. As more businesses ship MCPs, the ones with thoughtful UI responses — branded, theme-aware, action-oriented — will win the user's attention and trust inside the chat interface.
Aggregation MCPs will emerge. Just as I connected multiple flight services to plan a trip, users will expect their AI to combine MCPs from different providers. The businesses that play well together — standard formats, clean APIs, composable responses — will be the ones agents recommend.
The conversation becomes the transaction. MCP UI already supports interactive elements. We're not far from a world where you search for a deal, see the card, apply the coupon, and complete the purchase — all without leaving the AI chat. The entire customer journey happens in one conversation.
It's Not Perfect Yet.
For all its promise, MCP has real limitations that anyone building on it should understand.
The most frustrating one is tool selection: just because you've connected an MCP server doesn't mean the agent will use it. Unless you explicitly ask the agent to use a specific tool, it might default to web search, answer from memory, or ignore the MCP entirely. The model decides which tool to reach for, and if your server's tool descriptions aren't clear enough or if there are too many tools competing for attention the agent may simply skip yours.
GitHub Copilot found that cutting from 40 tools to 13 improved results significantly.
Then there's platform support. Gemini added MCP to its API and SDK in March 2026, but it's still limited to tools only, no resources, no prompts, and the consumer Gemini chat app still doesn't support connecting custom MCP servers the way Claude or ChatGPT do.
Cross-platform parity isn't here yet. On the UI side, MCP Apps render inside sandboxed iframes, and that sandbox is restrictive by design.
Links inside cards are just visual, clicking them opens in the sandbox context rather than navigating the user's browser normally, leading to broken pages or dead ends. The openLink API exists as a workaround (asking the host to handle navigation), but not every host implements it consistently, and Claude.ai currently has open bugs around CSP enforcement, it ignores declared frameDomains and connectDomains, hardcoding its own restrictive policy instead, which breaks any MCP App relying on external resources like product images from CDNs.
State persistence is another gap: the iframe can be destroyed and recreated at any point during a conversation, losing all in-memory state unless you build your own save/restore mechanism through server-side tools.
And the ecosystem is still young, host support for MCP Apps is confirmed in Claude Desktop, VS Code Copilot, Goose, Postman, and ChatGPT, but many other clients haven't implemented it yet. These are growing pains, not fatal flaws.
But anyone marketing MCP as plug-and-play today is overselling it.
For Businesses: Start Now
If your business has an API and a customer-facing product, you should be building an MCP server. Not because it's trendy because it's where your customers are going. The playbook is simple: take your core search and discovery functionality, wrap it in MCP, and invest in making the response visual and actionable.
At Enpitech, we built the Hi Mami MCP as a standard search-and-fetch server with rich UI responses. It's deployed on Vercel, connected to Claude as a native integration, and already listed on the MCP App Store. The technical complexity is manageable. The strategic value is enormous.
The AI assistant is becoming the new browser. MCP is how your business shows up in it.
Enpitech — Frontend Consulting for Startups. React · Next.js · TypeScript · AI Integration · MCP Development.
