Agent Experiences

MCP Server, Skill, Plugin, or App: What Your Product Actually Needs

By Alan WeibelPublished 11 min read

Someone has asked whether your product needs an MCP (Model Context Protocol) server. MCP is a shared standard that lets AI assistants such as Claude and ChatGPT use other software, and the server is the small piece you run so they can use yours. Maybe the question came from your CEO after a conference, a customer who wants to work with you without leaving their AI assistant, or a sales rep who saw a competitor announce an app. A week later someone else asks about skills, or plugins, or connectors.

Every AI company has its own word for plugging in: MCP servers, skills, plugins, apps, connectors, extensions, product feeds. It's hard to tell which of these are different things and which are one thing under another name, and harder still to tell which ones your customers will use. Build the wrong one and you can spend months on something few people open, but build nothing and you may be missing from the place where your customers have started to do their work.

The good news is that most of these names are packaging around just two things: what your software can do, and how to use it well. The packaging is the part that keeps changing. This post sorts the names into those two piles, shows what companies in fifteen industries built, asks four questions that tell you whether you need any of it, and ends with what to build first.

The short answer

Build one connection to what your software can do, and reuse it everywhere. For most products that's a remote MCP server (one you host, that customers connect to online) built on your API (the connection your own app already uses to talk to your systems). If you sell from a catalog, it may be a product feed (a file listing everything you sell). Next, write a skill that teaches agents (AI assistants that carry out tasks for someone, such as booking, buying, or filling in a form) how your customers work with you. Then add each platform's own listing only where your customers already spend their time. If people only come to you to read, you probably need none of it, and a website that agents can read will do.

Ten names, two things

NameMade byWhat it gives an agentWhere it runs
MCP serverAnthropic; now an open standardYour tools and dataYour server
Agent SkillAnthropic; now an open standardKnow-how and workflowsInside the agent
PluginEach agent vendorA bundle of the aboveInstalled in the agent
ChatGPT app / MCP AppOpenAI, AnthropicTools plus your own screensYour server, shown in chat
ConnectorAnthropic, Microsoft, GoogleA listing for your serverYour server
Custom GPT ActionOpenAIYour API (retiring)Your API
Product feedOpenAI, Google, MicrosoftYour catalogYour feed file
WebMCP (tools a web page offers to an AI agent in the browser)Early proposal from web standards groupsTools on your web pageThe user's browser
Agent2AgentGoogle; now an open standardLets your agent work with other companies' agentsYour agent
llms.txt (a list of key pages for AI tools) / NLWebCommunity; MicrosoftEasier-to-read contentYour site

Read down the "where it runs" column and a pattern shows up. An agent either calls something you run (a server, an API, or a feed), or it carries instructions you wrote (a skill). Plugins, apps, and connectors are the listings and bundles each assistant uses to find those two things.

  1. Where your customers already are

    Each assistant has its own listing and its own packaging. Underneath, they all wrap the same thing, so you don't build each one from scratch.

    • ChatGPT plugin or app
    • Claude connector or plugin
    • Microsoft 365 Copilot agent
    • Gemini
    • Cursor and coding agents
  2. Your capabilities

    You build this part once. It does the real work, and every wrapper above reuses it.

    • Remote server using MCP (a shared standard that lets AI assistants use your software)
    • A public API (the way other software connects to a product directly)
    • A product feed (a file listing your products and prices) and a way for agents to check out
  3. How to use them well

    Written instructions the agent reads when a task calls for them. They are cheap to write and easy to change.

    • Agent Skill (SKILL.md)
  4. Underneath it all

    What an agent falls back on when nothing above exists.

    • Your website, with its text in the page from the start, not added later by scripts
    • WebMCP (tools your page offers to an agent in the browser)
    • Structured data
Most of the names in the top box are packaging for the “Your capabilities” layer.

What each one is

MCP server

The Model Context Protocol is a shared standard for describing what your service can do in a way any AI agent can use. An MCP server is the piece you run: an agent asks it what it can do, the server lists its tools (the actions it offers, such as checking stock or creating an invoice), and the agent calls them when a task needs one. Anthropic released the protocol in November 2024 and gave it to the Linux Foundation's Agentic AI Foundation in December 2025. ChatGPT, Claude, Gemini, Microsoft Copilot, Cursor, and VS Code (two tools developers write code in) all support it. The MCP project reports that its main SDKs (the code libraries developers use to build servers and clients) are downloaded close to half a billion times a month.

There are two kinds. A local server runs on the user's own computer, which suits developer tools. A remote server runs on your company's servers and signs users in with OAuth (the familiar "sign in and allow access" step, which lets someone connect an account without handing over their password). The remote kind is the one that matters for most products. The July 28, 2026 update made remote servers simpler to host, like any other web service. Our implementation patterns page covers how to design the tools.

Agent Skill

A skill is a folder with a SKILL.md file in it: a name, a description, and written instructions, plus any scripts or templates the task needs. At first the agent reads only the name and description, and opens the rest when a task needs it. Agent Skills started at Anthropic in October 2025 and joined the Agentic AI Foundation in October 2026. OpenAI says its skills build on the open standard, and GitHub Copilot added support in December 2025.

A skill mostly teaches an agent how to use tools it already has: which step comes first, which settings your customers always want, and what to check before anything is sent. Of everything on this list, it's the cheapest to write and the easiest to change.

Plugin

A plugin is a bundle of skills, MCP servers, commands, and sometimes hooks (small scripts that run automatically at set moments), all installed together. Claude plugins install from marketplaces, and a marketplace can be any git repository (a shared folder of code, such as one on GitHub) with a listing file in it. OpenAI's plugins work across ChatGPT and its coding agent, Codex, from one catalog. Cursor opened its marketplace in February 2026.

So a plugin is a way of getting what you built in front of people. Make one once you have a server and a skill to put in it.

ChatGPT apps and MCP Apps

Some results only make sense when you can see them, like a map of houses for sale, a seating plan, or a design. OpenAI's Apps SDK, launched in October 2025, lets your MCP server send back screens you designed, to be shown right inside the chat. In January 2026, maintainers from OpenAI and Anthropic published MCP Apps as a shared extension, so one set of screens can appear in Claude, ChatGPT, VS Code, and Microsoft 365 Copilot.

Connectors

A connector is mostly a directory listing for a remote MCP server, so that people can find it and switch it on. Claude's connectors directory listed 1,094 of them on November 8, 2026. Microsoft 365 Copilot has two kinds of connector. Synced ones copy your content into Microsoft Graph (where Microsoft 365 keeps a company's work data), and federated ones ask your service live, over MCP, each time. Gemini in Google Workspace connected to outside tools over MCP in September 2026.

Custom GPTs and GPT Actions

A GPT Action described your API in OpenAPI (a standard format for writing down what an API can do), so that one custom GPT (a version of ChatGPT someone set up for one job) could use it. It only ever worked in one product at one company, and that product is being retired. Custom GPTs stop working on December 11, 2026, and OpenAI's retirement FAQ says custom actions don't carry over. If you have one, rebuild it as a remote MCP server, and it will work with every assistant here that supports MCP, once your customers connect it.

Product feeds

If you sell physical products, start with your catalog. A product feed is a file listing what you sell, with prices and stock, that AI shopping services read. ChatGPT, Google, and Microsoft Copilot all take them. Profound, which tracks ChatGPT Shopping results, found that on July 10, 2026, the share of ChatGPT Shopping recommendations coming from merchants with a product feed jumped from about 8% to about 62%. Search Engine Journal also reported these figures. We dug into those numbers in our analysis of ChatGPT shopping.

Buying inside the chat has had a harder time. OpenAI moved purchases back to merchants in March 2026, though the Agentic Commerce Protocol (a shared standard for letting an agent pay on someone's behalf) that it built with Stripe carries on. Our product feeds guide covers the fields, the platforms, and the mistakes that get products dropped.

WebMCP

WebMCP lets a web page offer tools to an AI agent running in the user's browser, using the account they're signed in to. Chrome began an origin trial (an early test that websites sign up for before a feature reaches everyone) in version 149, and Edge is running one too. It's a low-effort way to let an agent use a complicated page properly, and it's still experimental.

Agent2Agent

Agent2Agent is for when you have your own agent and want it to work with someone else's, handing over a task and getting a result back. It reached version 1.0 in March 2026 and joined the Agentic AI Foundation in August. Most products don't have an agent of their own yet, so most don't need it.

llms.txt and NLWeb

These two help agents read your content. llms.txt is a plain text file on your site that lists your most important pages for AI tools, and NLWeb is a Microsoft project that lets a site answer questions about its own content in everyday language. Neither one lets an agent do anything on your site. We looked at whether anything fetches llms.txt in The llms.txt Verdict, and the short version is that very little does.

Who built what

IndustryCompanyWhat they built
TravelExpedia, Booking.comChatGPT apps
Real estateZillowChatGPT app with maps
MusicSpotifyChatGPT app
EducationCourseraChatGPT app
DesignFigma, CanvaChatGPT apps and remote MCP servers
RetailEtsy, Target, WalmartProduct feeds, then ChatGPT apps
GroceryInstacartChatGPT app, then a Claude connector
PaymentsStripe, PayPalMCP servers, commerce protocols
BankingGrasshopper BankRead-only MCP server
FinanceIntuitChatGPT apps for TurboTax and others
HealthcareHealthEx, Function HealthClaude connectors for records and labs
LegalLegalZoomClaude connector and plugin
DomainsGoDaddyPublic, read-only MCP server
Work toolsAtlassian, LinearRemote MCP servers
Dev toolsVercel, SupabaseMCP servers and plugins

What each company chose follows what its product does, not which industry it's in.

Zillow, Expedia, and Figma each give back something you have to look at, so they built apps. A list of 40 houses as text isn't much help when a map is right there, and these companies built screens to show it inside the chat.

Atlassian, Linear, Stripe, Vercel, and Supabase went for servers instead. They sell to developers and office teams who use AI assistants, and those customers want an agent to create the ticket or change the setting, which plain text handles fine. Retailers took a third route: Etsy, Target, and Walmart were part of AI shopping through product feeds and checkout partnerships before any of them shipped an app.

In regulated industries, the scope stayed narrow. Grasshopper Bank's server is read-only, so an agent can look at balances and transactions but can't move money. Health records reach Claude through record platforms like HealthEx, rather than through each hospital. LegalZoom's connector hands the conversation to an attorney instead of answering the legal question itself.

GoDaddy started with the smallest useful piece. Its server checks whether a domain name is available, with no sign-in needed. That's a single read-only task, but it means an agent with GoDaddy's server connected can answer straight from GoDaddy when someone asks whether a name is free.

Do you even need one?

Not every product does. Ask these in order.

  1. 1Do people come to you only to read: articles, docs, prices, opening hours?

    Yes
    You don't need to build a server. Make sure agents can read your pages, and keep your structured data (labels in your code that state facts such as the author, date, or price) accurate.
    No
    Next question.
  2. 2Do you sell products from a catalog?

    Yes
    Start by sending a product feed (a file listing your products and prices) to ChatGPT, Google, Microsoft, and Perplexity. A server can come later.
    No
    Next question.
  3. 3Do people take actions in your product: book, file, create, change, buy a service?

    Yes
    Build a remote MCP server that people sign in to, then write a skill (a set of instructions for the agent) for the tasks people do most.
    No
    Your website is all an agent needs. Keep it easy for an agent to read and to use.
  4. 4Does the result only make sense visually: a map, a design, a seat plan, a playlist?

    Yes
    Add an app (ChatGPT app or MCP App) on top of the same server.
    No
    Stay with the server and the skill. Text results are enough.
Do you need one? Stop at the first yes, except the last question, which adds to the one before it.

A news site, a law firm's practice pages, or a restaurant's menu gains nothing from an MCP server. What an agent needs from them is a page it can fetch and understand, and the Agent Analyzer checks exactly that. A retailer gets further with a clean product feed than with a custom app. The products where a server pays for itself are the ones where people book, file, create, or change things, because without a server an agent has to drive your interface like a person, clicking through slowly, with a good chance of failing partway.

The trade-offs

OptionBuild effortReachMain risk
Remote MCP serverMediumEvery major assistantSecurity and sign-in
Agent SkillLowMost agentsGoes stale as you change things
PluginLow, given the partsOne ecosystem eachEach store's review rules
App (ChatGPT or MCP Apps)HighChatGPT, Claude, CopilotLooks different in each assistant
Connector listingLowOne directory eachAdmins must switch it on
GPT ActionDo not start oneChatGPT until Dec 11Retiring
Product feedLow to mediumAI shopping servicesFeed and page disagree
WebMCPLowChrome and Edge trialsExperimental
Agent2AgentHighMostly large companiesStill early

Two warnings apply to every row.

First, an integration that exists may not get used. Digiday reported almost 900 apps in ChatGPT by May 2026, and said it was unclear whether shoppers used them. A listing in a directory doesn't put you in the conversation, and nobody publishes how assistants choose what to suggest, so make yours the most useful for the tasks your customers bring.

Second, these formats get retired. ChatGPT plugins lasted about a year before they stopped working in April 2024. Custom GPTs lasted three, and the thing replacing them is called plugins again. Each time, companies that had built straight onto the packaging had to start over, while companies whose packaging sat on top of an API kept the API and rewrote a thin layer. MCP is the safest base so far, because no single company owns it.

What to build, in order

  1. Make sure agents can read your website. Have your pages arrive with the content and prices already in them, rather than filled in later by scripts, and keep your structured data (labels in your page's code that state facts such as a product's name, price, or reviews) accurate. Every agent falls back on your pages, and every click from an assistant lands on one.
  2. If you sell products, publish a feed. Follow the product feeds guide, and keep the feed and the page in agreement.
  3. If people take actions with you, build a remote MCP server. Put it on the API your own app already uses, so an agent can do whatever a customer can. We wrote about what happens when your app can do something your API can't in No Capability Behind a UI. Offer a few tools shaped around real tasks rather than one for every endpoint (each address in your API that software can call), as the implementation patterns page explains.
  4. Write a skill. Describe the workflows your customers run, in the order they run them. It's a few hours of work, and it makes the server much more useful.
  5. List it where your customers are. That means a Claude connector and a ChatGPT plugin if they live in chat, a Copilot connector if they live in Microsoft 365, and a Cursor or Claude Code plugin if they're developers. Add an app only if your results need to be seen.
  6. If you have a GPT Action, move it before December 11. Rebuild it as in step 3, and it will work with every assistant here that supports MCP, once your customers connect it.

This order is also the order of what lasts. Your website and API will probably outlast every name in the first table. They are what engineers placed inside customer companies have to reach before an agent can do any real work, and what custom software made of AI agents is built on.

What this does not tell you

The counts here are snapshots: Claude's directory on November 8, 2026, and Digiday's report from May 2026. Both change every week. Nobody publishes how often each kind of integration is used, so "reach" in the trade-offs table means "can be installed", not "is used". The names will keep changing too. Google is moving Gemini's Gems (its custom assistants) to something it calls skills from November 17, 2026, and it hasn't said whether these follow the open SKILL.md standard. Getting found by AI search has a similar naming muddle, with people calling it GEO (generative engine optimization), AEO (answer engine optimization), and more. Our advice there also holds here, which is to ignore the label and check what the thing does. We wrote about it in GEO, AEO, LLMO: What Do You Actually Call This?.