What is an MCP server? A plain-English guide

Your AI assistant can draft a follow-up email in seconds. Ask it who registered for last week's webinar, though, and it has nothing to go on. An MCP server closes that gap. It's a small program that connects an AI assistant like Claude or ChatGPT to one specific tool or data source, using an open standard called the Model Context Protocol (MCP). The server tells the assistant what it can read and do in that tool, then does it when you ask.

You don't need to write code to use one. Most MCP servers you'll meet as a marketer come ready-made from software you already pay for, and connecting one takes about as long as logging in.

What is an MCP server?

An MCP server is a program that gives an AI application access to a tool or data source through the Model Context Protocol. The protocol is the shared language. The server is the piece that sits next to your tool and speaks that language on its behalf.

What MCP stands for

MCP stands for Model Context Protocol. The official project describes it as "an open-source standard for connecting AI applications to external systems" (modelcontextprotocol.io). "Context" is the word that matters most. It's the information an AI model needs to give you a useful answer, like your files, your calendar or your webinar registrations.

The project's own analogy holds up well. Think of MCP as a USB-C port for AI applications, one standard connector instead of a drawer full of adapters.

Anthropic, the company behind Claude, open-sourced MCP in November 2024 as an open standard anyone can build on.

The "server" part, in plain English

The word "server" sounds bigger than it is. In MCP's own definition, a server is "a program that provides context to MCP clients." Nothing in that definition says it has to be big.

Each server usually covers one tool. Your calendar can have one, your document storage another, your webinar platform a third. The server describes what an AI can do in that tool (look up a record, create an item, pull a list) and carries out the request when the AI asks.

A one-minute example

Say you ask your AI assistant: "Which accounts registered for Tuesday's webinar but haven't come to a session this quarter?"

Without an MCP server, the assistant can only guess, or ask you to paste in a spreadsheet. With your webinar platform's MCP server connected, it looks up the registrations and attendance itself and answers from real data. You skip the CSV export and the spreadsheet gymnastics.

Why MCP servers exist

Before MCP, connecting an AI assistant to a business tool meant building a custom integration for every assistant and every tool, one pair at a time.

The integration problem before MCP

Picture five AI assistants and twenty business tools. Wiring all of them together one by one means up to a hundred separate integrations, each with its own quirks and its own upkeep.

Anthropic's announcement noted that every new data source required its own custom implementation. MCP replaces those "fragmented integrations with a single protocol," in Anthropic's words. A software company builds one MCP server, and any assistant that supports MCP can use it. An AI company adds MCP support once, and every MCP server becomes available to its users.

What changes for business teams

The official MCP documentation describes the payoff for end users plainly: AI applications that "can access user data and take actions on the user's behalf when necessary."

That's the real shift. An assistant with no access to your tools can only talk about your work. Connected through MCP servers, the same assistant can look things up and get tasks done inside those tools, within the permissions you give it.

Why marketers should pay attention now

Marketing teams already use AI, but far less inside the tools where campaigns run. Our 2026 Webinar Benchmark Report shows webinar-specific AI adoption (59.8%) trailing general marketing AI use (80.2%) by 20.4 points. The same report puts average webinar-AI maturity at 2.54 out of 5, with roughly half of the concerns tied to trust and accuracy.

That trust gap is where MCP earns its place. In Livestorm's October 15 webinar, Building an agentic webinar system, Tata Maytesyan, who builds AI marketing systems, put it bluntly:

An assistant that reads your actual registration list through an MCP server is answering from your data instead of guessing.

How an MCP server works

Three pieces work together, and you already know the first one.

Hosts, clients and servers

The host is the AI application you use: Claude, ChatGPT, or a code editor like Visual Studio Code. MCP's documentation defines it as "the AI application that coordinates and manages one or multiple MCP clients" (MCP architecture overview).

Inside the host, a client keeps a dedicated connection open to one MCP server. Connect three servers, and the host runs three clients, one for each.

The server sits on the tool's side. It answers the client's questions about what it can do, and it does the work when asked.

Tools, resources and prompts

An MCP server can offer three kinds of things, which the protocol calls primitives.

• Tools are actions the AI can take. MCP calls them "executable functions," and a webinar example would be registering an attendee or creating an event.

• Resources are data that gives the AI context, such as the attendee list from your last session.

• Prompts are reusable templates that structure a task, like a standard post-webinar recap.

Tools get most of the attention because they're where the visible work happens. Resources and prompts are quieter, but they shape what the assistant knows and how it starts a task.

What happens when you ask a question

Here's the sequence behind a single request:

1. When you connect, the AI app asks each MCP server what it offers, for example by requesting its list of tools.

2. You ask a question in plain English.

3. The model decides which tool fits, and with what details (which webinar, which date).

4. The MCP server runs that action in the tool, using the permissions you granted.

5. The result comes back into the conversation, and the assistant answers from it.

Your part is step 2. The rest happens behind the chat window.

Local vs remote MCP servers

MCP servers come in two kinds, and the difference mostly decides how you connect. A local MCP server runs on your own computer. A remote MCP server runs on the provider's infrastructure, and you connect to it over the internet, usually by signing in.

Local servers

A local server runs on the same machine as your AI app and talks to it directly through what MCP calls the STDIO transport. The official example is the filesystem server, which Claude Desktop launches on your computer so Claude can work with your local files.

Local servers are mostly a developer thing. They take some setup, and because they run code on your machine, you need to trust where they came from.

Remote servers

A remote server runs on the provider's platform and uses the Streamable HTTP transport. The official example is Sentry's MCP server, which runs on the Sentry platform. A single remote server typically serves many users at once.

MCP recommends OAuth for authentication, the same standard behind most "Sign in with..." buttons. That's why connecting a remote server usually feels like logging in to an app: you approve access, and you're done.

If you work in marketing, sales or operations, remote servers are the ones you'll meet.

MCP server vs API (and vs RAG)

These three terms get mixed up a lot, and the differences are easy to pin down.

MCP server vs API

An API is the door a piece of software opens to other software. Developers write code against it, call specific endpoints and handle the results. It's precise, and it needs a developer.

An MCP server usually sits on top of that API. It describes the available actions in a format an AI assistant can discover on its own, so the assistant can ask "what can you do?" and call the right action when your question needs it.

Livestorm is a good example. Its MCP server exposes the Livestorm Public API as tools that AI assistants can use.

MCP vs RAG

RAG, short for retrieval-augmented generation, pulls relevant information into an AI model's prompt so its answer is grounded in your content. Google Cloud's guide draws the line clearly: "RAG finds and uses information for creating text, while MCP is a wider system for interaction and action" (Google Cloud).

The two work well together. RAG improves what the model knows when it answers. MCP standardizes how the assistant connects to tools, whether it's reading data or taking an action like registering the leads it found.

Is an MCP server a real server?

Yes, though not in the rack-of-machines sense. The MCP documentation says the term refers to the program that serves context data, regardless of where it runs. That can be a small process on your laptop or a hosted service in the cloud.

Which AI tools support MCP?

MCP is supported across AI assistants and developer tools. The official MCP site lists Claude and ChatGPT among the assistants, and Visual Studio Code and Cursor among the developer tools.

AI assistants

In both Claude and ChatGPT, you add an MCP server as a connector in the settings, then ask questions that use the connected tool. So if you've been wondering whether ChatGPT uses MCP, it does.

Developer tools

Developers got there first. When Anthropic announced MCP, it named developer-tool makers such as Zed, Replit, Codeium and Sourcegraph among the companies working with the protocol. Editors like Visual Studio Code and Cursor now let a coding assistant reach a project's files and connected services through MCP servers.

Connectors and directories

For non-developers, the bigger change is the directory. Claude has a directory of connectors, so adding an MCP server can be as simple as finding the app and clicking Connect, with no URL to paste and no settings file to edit.

Are MCP servers safe? What to check before you connect one

An MCP server is as safe as the company behind it and the permissions you give it. In Livestorm's October 15 webinar, Tata Maytesyan was candid about weighing that risk:

That's the same logic you'd apply to any app you let into your work accounts.

The real risks, in plain terms

The official MCP security best practices spell out the risks worth knowing. Local servers are programs that run on your machine, so one from an untrusted source can run commands with the same access as your AI app. Broad permissions are the other common problem: if an access token with sweeping rights leaks, everything those rights cover is exposed. The guidance favors starting with minimal permissions and adding more only when a task needs them.

Consent matters as well. Google Cloud's guide lists user consent and control first among its MCP security principles: you should understand, and approve, what an assistant can read and do through a server.

A five-point checklist before you connect

1. Who built it? Prefer the official server from the tool's own vendor over an unofficial copy.

2. How does it sign you in? An OAuth sign-in is safer than pasting API keys into a config file.

3. What can it access and change? Check whether it only reads data or can also create, edit and delete.

4. Whose permissions does it use? The assistant should be able to do what your own account can do, and nothing more.

5. How do you disconnect? You should be able to revoke access from the AI app's settings at any time.

None of this needs a security background. It's a short check, and it's the same one your IT team will ask you about anyway.

What an MCP server looks like for a webinar platform

Here's the concrete version. Livestorm, the webinar platform with AI Studio, runs an official remote MCP server. Connect it to Claude, ChatGPT or Microsoft 365 Copilot, sign in with your Livestorm account, and you can ask in plain English to create a webinar, register a list of contacts, or see who attended but didn't convert.

How it connects

The Livestorm MCP server lives at https://mcp.livestorm.co/mcp. It's a remote server that uses Streamable HTTP and OAuth 2.1 with dynamic client registration, so AI apps register themselves and you sign in through a normal Livestorm login screen.

In Claude, Livestorm is a verified connector in Claude's directory, which means Anthropic reviewed it for quality and compatibility. It was listed in August 2026, and you can connect it from Claude's directory in one click. It works in Claude on the web and desktop, and in Claude Code.

It also works with ChatGPT as a custom connector, Microsoft 365 Copilot through Copilot Studio, GitHub Copilot in VS Code, Cursor, Gemini CLI and Dust. Other MCP clients connect through the mcp-remote bridge, which is the only setup that needs Node.js.

What you can ask it to do

In Livestorm's October 15 webinar, Head of Growth Benoit Leggieri summed up what the server does:

You can list, create, update and delete events and sessions, register attendees one by one or in bulk, look up and tag people, and read attendance, chat, Q&A and recordings. Team-member management and webhooks are included too.

In practice, the requests look like this:

• "Create a webinar next Tuesday at 3pm and register these contacts."

• "List my webinars this month."

• "Who attended my last webinar but did not convert?"

That last question usually means an export, a spreadsheet and a CRM lookup before anyone gets an answer. One note on CRM sync: it runs through webhooks or your CRM's own connector, so keep that integration in place alongside the MCP server.

Livestorm's October 15 webinar was set up through this server. On October 6, Tata Maytesyan asked Claude to create the event, invite Benoit Leggieri as co-host and pull stats from her past webinars. The event page was live in under five minutes, and Claude flagged two fixes for a human: a 30-minute session instead of 45, and a custom registration question the MCP couldn't add. She then built an engagement report on the same connection, with signups by source, company, region and industry.

Who controls access

The connection uses OAuth with the same permissions as the Livestorm user who connects it, so the assistant can't see or change anything that person couldn't. Workspace owners and admins connect it, it's available on every Livestorm plan including Free, and you can disconnect it at any time.

The Livestorm MCP help article walks through setup for each AI app, and our post on how Livestorm's MCP server makes webinar planning easier covers more use cases.

Do you need an MCP server? How to get started

If you're a marketer or an ops lead, you don't build an MCP server. You connect the ones your tools already offer. Developers build servers for their products and internal systems; everyone else signs in and starts asking questions.

Build one or connect one?

Building makes sense when you ship a product, or run an internal system, that other people's AI assistants should be able to use. That's a developer project, usually built on top of an existing API.

Connecting is everyone else's job, and it's where the value shows up first. If the software you use every day already has an official MCP server, you're one sign-in away from asking your assistant about it.

Get started in four steps

1. Pick the AI assistant your team already uses, such as Claude, ChatGPT or Microsoft 365 Copilot.

2. Find your tool's connector in the assistant's directory, or add the server URL the vendor provides.

3. Sign in, and read the permissions screen before you approve.

4. Start with a read-only question, like "List my webinars this month," before you ask it to create or change anything.

Start with the tool that eats the most of your week. For most webinar teams, that's the platform where registrations and attendance live. To try it on your own webinar data, create a free Livestorm account and connect the MCP server from your AI assistant.