Why SaaS and apps need an MCP server to get found in ChatGPT and Claude

Your next customer may never visit your website. They will ask ChatGPT or Claude for a tool that does the job, and the assistant will pick one. If it can use your product from inside the chat, you are the easy pick.

That is what an MCP server does. Here is what to build, what it takes, and when it is too early.

What is an MCP server, in plain words?

MCP, the Model Context Protocol, is an open standard that lets an AI assistant call outside tools. An MCP server is the piece you build. It tells the assistant what your product can do, such as “create a report”, “check availability” or “send a quote”, and does it when the user asks.

The user stays in the chat. They type what they want, the assistant calls your server, and the result comes back in the conversation. They sign in to your product once, the first time.

Anthropic introduced MCP. In December 2025 it handed MCP to the Agentic AI Foundation, a fund under the Linux Foundation that Anthropic co-founded with Block and OpenAI. By then ChatGPT, Gemini, Microsoft Copilot, Cursor and VS Code all supported it, and there were more than 10,000 active public MCP servers.

Why does it help new people find you?

Because buyers have moved their first question. In G2’s April 2026 survey of 1,076 software buyers, 51% now start their research with an AI chatbot more often than with Google, up from 29% a year earlier. 69% had picked a different vendor from the one they planned on, because the chatbot recommended it.

The assistants now have shop windows. ChatGPT has an app directory, open to submissions since December 2025, and every app in it is built on MCP.

OpenAI now calls them plugins, and in September 2026 it said ChatGPT recommends relevant plugins inside conversations, not only in the directory. ChatGPT has 1.2 billion weekly users, by OpenAI’s own count at the same event.

Claude has a connectors directory, open since July 2025. Each connector is an MCP server. Anthropic said it had grown past 200 connectors by April 2026, available on every Claude plan.

Is being cited by ChatGPT enough?

No. A citation means the assistant mentions you and maybe links to you. The user still has to leave, find your site, sign up and learn your product.

With an MCP server, the assistant can do the work with your product while the user watches. The first thing the user sees is a result, not a sign-up form.

Citations still matter, and we build for both (see how Shhots AI grew its new users from ChatGPT by 684%). But a product the assistant can use is harder to swap out than a link.

Which jobs should your MCP server do?

Pick the few jobs people would ask an assistant for in one sentence. Don’t expose every endpoint in your API. The assistant reads each tool’s description to decide what to call, so a short list of clear jobs beats a long list of vague ones.

Business	Jobs worth exposing
Analytics or reporting SaaS	Pull this month’s numbers, compare two periods, explain a change
Design or content app	Create an asset from a brief, resize it, export it
Scheduling or booking product	Check availability, book a slot, reschedule
Agency or service business	Describe what you offer, check if a lead fits, book a call, send a quote
Ecommerce or marketplace	Search the catalogue, check stock and delivery, track an order

Service businesses belong on that list. An assistant that can check whether a lead fits your criteria and book the call does more for you than one that links to your contact page.

What does it take to build one?

Part	What it involves
The jobs	Choosing which jobs to expose, and writing a description of each that the model reads
Sign-in	OAuth, so each user connects their own account and the assistant acts only on their behalf
The server	A remote MCP server, hosted and monitored, that calls your product’s API
Listings	Submitting to the ChatGPT directory and the Claude connectors directory, each with its own review
A page about it	A page on your site that explains the server, so search engines and assistants can find it
Measurement	Connections, calls per job and sign-ups that came through the assistant

The specs move. OpenAI renamed apps to plugins within a year, and both directories have changed their rules since launch. Someone has to keep up after it ships.

Who should wait?

• Products with no API yet. The MCP server calls your product. Build the API first.
• Products whose job doesn’t fit one sentence. If a user can’t ask for the result in plain words, the assistant can’t either.
• Categories nobody asks assistants about. Run your buyers’ questions through ChatGPT and Claude first. If nobody asks, a citation plan comes first.

What happened when Shhots AI built one?

Rajat built an MCP server for Shhots AI, his AI ad generator, in August 2026. Users connect it to Claude or ChatGPT and ask for a UGC video ad, a product photoshoot or a video from one product photo. The finished ad comes back in the chat.

It is included in the paid plans at no extra cost, and it uses the same credits as the app.

In our 27 Oct 2026 check, ChatGPT named Shhots second when asked for an MCP server that makes video ads, and cited the launch post. The MCP server put Shhots into a new kind of answer: which tool can make the ad from inside the chat.

How does XYZ Lab build it for you?

The same way we do everything else: Diagnose, Build, Run.

• Diagnose. We run your buyers’ questions through ChatGPT and Claude, see who gets named, and pick the jobs worth exposing.
• Build. We build the server in your accounts, set up sign-in, write the tool descriptions, and submit to both directories.
• Run. Each month you see connections, calls and sign-ups through the assistants. When OpenAI or Anthropic change the rules, we update the server.

It is part of AI for product and AI for growth. To find out whether your product should be inside ChatGPT and Claude, pick a time in the calendar below.