ArticleAI Workflows

What Is MCP (Model Context Protocol)? A Guide for Marketers

MCP (Model Context Protocol) is an open standard that connects AI assistants to your tools and data. How it works, what it changes for marketers and how to use it safely.

Published 18 min read
02 / In this article

MCP, short for Model Context Protocol, is an open standard that lets an AI assistant connect to outside systems such as your analytics, your ad accounts or your files, using one shared set of rules instead of a custom integration for each pair. For a marketer, it is the layer that lets you ask an assistant a question about your own data and get an answer based on the real numbers. This guide explains the problem MCP solves, how the pieces fit together, what it changes in daily marketing work, what can go wrong and what it is not.

Key Takeaways

  • MCP is a connection standard, not an AI model and not a product. It defines how an AI application talks to a server that exposes data and actions.
  • Three roles matter: the host (the AI application you use), the client (the connector inside it) and the server (the program that exposes your data or tools).
  • Servers offer three kinds of building blocks: tools the model can call, resources it can read and prompts a person can start.
  • For marketers, the gain is access to live data in a conversation: search performance, ad results and analytics questions answered from the source rather than from a pasted export.
  • Start read-only. The specification says there should be a human in the loop who can deny tool calls, and the official security guidance recommends least-privilege access.

What is MCP (Model Context Protocol)?

MCP is an open-source standard for connecting AI applications to external systems. The official documentation describes it with an analogy: think of MCP like a USB-C port for AI applications. A USB-C port gives many devices one standard way to connect, and MCP gives AI applications one standard way to reach data sources, tools and workflows.

Anthropic introduced the protocol on 25 November 2024, together with the specification, software development kits, support for local servers in the Claude desktop apps and an open repository of example servers. Since then it has been adopted well beyond that first release. The project's own introduction says AI assistants such as Claude and ChatGPT, and development tools such as Visual Studio Code and Cursor, support it. As of October 2026 the current version of the specification is dated 2026-07-28.

Two points keep the term clear. First, MCP is about the connection, not the intelligence. The documentation states that MCP "focuses solely on the protocol for context exchange" and does not dictate how an AI application uses language models or manages the context it receives. Second, "MCP" is the standard and "MCP server" is one program that follows it. When people say they "installed an MCP", they usually mean they connected an MCP server to an assistant.

What problem does MCP solve?

MCP solves the problem of connecting every AI application to every data source one custom build at a time. Anthropic's announcement describes it directly: each new data source needed its own custom implementation, which made connected systems hard to scale. The protocol replaces those separate connectors with one standard that both sides can follow.

Think about what this looked like for a marketing team. You want an assistant to answer "which pages lost clicks last month?" The data sits in a search tool. The assistant sits somewhere else. Without a standard, someone has to build a bridge for that assistant and that tool, then repeat the work for the next assistant, the next tool and every change in either one. With a standard, the tool vendor or a developer builds one server, and any assistant that supports the standard can use it.

BEFOREWithout a standardEach assistant needs its ownbridgeEach tool needs its own bridgeEvery change breaks a connectionData pasted by hand as exportsAFTERWith MCPOne server per data sourceAny supporting assistant can useitCapabilities are listedautomaticallyQuestions run against live data
One standard replaces a custom bridge for every pair.

The benefit is not only fewer builds, and it is the reason AI workflows built around your own data have become practical to scope. The documentation lists who gains: developers spend less time on integrations, AI applications gain access to a wider set of data sources and tools, and end users get assistants that can reach their data and, when they allow it, act for them. That last clause matters for marketers, and we come back to it in the safety section.

How does MCP work?

MCP follows a client-server design. An AI application, called the host, connects to one or more servers by creating one client for each server. Each client keeps a dedicated connection to its server, and the server provides context and capabilities to the host.

The three roles are easy to place with an example from the official documentation. Visual Studio Code acts as a host. When it connects to a server, it creates a client object to keep that connection alive. When it connects to a second server, it creates a second client. In your case, the host is the assistant you chat with, the server is the connector to a tool such as your search or analytics data, and the client is the part inside the assistant that holds the line open.

  • MCP host: the AI application that coordinates one or several clients.
  • MCP client: the component that maintains a connection to one server and gets context from it for the host.
  • MCP server: the program that provides context. It can run on your own computer (a local server) or on a vendor's platform (a remote server).

The protocol itself has two layers. The data layer defines the messages, built on JSON-RPC 2.0, including how a client discovers what a server can do. The transport layer defines how messages travel. MCP supports two transports. The stdio transport connects local processes on the same machine. The Streamable HTTP transport uses HTTP for remote servers and supports standard authentication, and the documentation recommends OAuth to obtain tokens. You do not need to handle any of this to use MCP, but it explains why some servers run on your laptop and others are hosted.

Want to connect an assistant to your marketing data?

Tell us which questions you ask your reports every week. We map which parts could run as an AI workflow and which should stay manual.

Explore AI workflows

What can an MCP server offer: tools, resources and prompts?

An MCP server can expose three building blocks, called primitives: tools, resources and prompts. Each is controlled by a different party, and knowing who controls what tells you where the risk sits.

PrimitiveWhat it isMarketing exampleWho controls it
ToolsFunctions the model can call, which can read or change thingsRun a query on search data, pause a campaignThe model
ResourcesRead-only data that supplies contextA report, a schema, a documentThe application
PromptsReusable instruction templates"Summarize last week's performance"The user

Tools are the powerful part. The documentation defines them as executable functions that an AI application can invoke to perform actions, such as file operations, API calls or database queries. The model decides when to call one, based on your request. A tool can write as well as read, which is why permission settings matter more here than anywhere else.

Resources provide data for context. Examples in the documentation include file contents, database records and API responses. They are application-driven: the host decides how to retrieve and present them.

Prompts are templates that a person starts on purpose. Typical forms are slash commands or entries in a command palette. A server author might include a prompt that walks the model through a standard weekly review of a data source.

The same server can offer all three. The documentation's own example is a database server: it can expose tools for querying the database, a resource holding the schema, and a prompt with examples of how to use the tools. A client discovers what is available by asking the server to list its tools, resources and prompts, and a client can ask again when the list changes.

What happens when you ask an assistant a question with MCP?

When you ask a question, the assistant decides whether a connected tool would help, calls that tool through its MCP client, receives the result and writes an answer from it. The model never connects to your data directly. The host and its client sit in between and carry out the call.

Take a plain question: "Which landing pages had the biggest drop in clicks last month, and what changed on them?" The flow runs in five steps.

  1. You ask the question in the assistant.
  2. The assistant sees the connected tools. Earlier, the client asked each server what it offers, so the model knows a search-data tool exists and what inputs it needs.
  3. The model picks a tool and fills in the inputs, such as the date range and the metric.
  4. The client sends the call to the server, and the server fetches the data from the underlying system using the permissions it has been given.
  5. The result returns to the model, which turns it into a readable answer, often with a table or a short explanation.
01QuestionYou askin plainwords02PicktoolModelchoosesfrom thelist03ClientcallHostasks theserver04ServerfetchReadsfromyouraccount05AnswerModelwritesit up
From a marketing question to an answer from live data.

One detail is worth knowing: the protocol does not decide whether a call needs your approval. That is a feature of the host. Good hosts show you which tool is about to run, with which inputs, and ask before anything sensitive happens.

What does MCP change for a marketer?

MCP lets a marketer ask questions of live data in everyday language, in the same place where the thinking happens, instead of exporting a file, pasting it and hoping the assistant reads it correctly. It does not replace the analysis. It shortens the path between a question and the data.

The most useful changes are practical:

  • Search performance. Ask which queries or pages gained or lost impressions and clicks, and compare two periods, from the source data rather than a screenshot.
  • Paid media. Ask how campaigns, ad groups or ads performed over a period and which ones deserve a closer look.
  • Analytics. Ask which channels, pages or events moved, and whether a change lines up with a launch or a campaign.
  • Several sources at once. Because one host can hold many connections, a single question can draw on search data and analytics data together.

This is the idea behind the dedicated guides we are publishing on each source: connecting Google Search Console to an assistant, working with Google Ads data and querying Google Analytics. Each one covers what exists for that product and how to check the source before you trust it. If you want a wider map before choosing, our overview of MCP tools for marketers compares the options by job.

The limit is that MCP gives an assistant access, not judgment. You still decide which numbers matter and what a change means. A team without a measurement plan will get fast answers to poorly framed questions. If your setup needs that groundwork first, our analytics and tracking service covers the measurement layer an assistant would read from.

Good fits are repetitive, checkable jobs: a weekly summary or a list of pages with falling clicks. Poor fits are jobs where a wrong answer is costly or hard to spot, such as changing bids, budgets or tracking without review. Ask of any task: if the assistant is wrong, who notices, and how quickly?

Start hereRead, low impactRead with reviewRead, high impactDraft onlyWrite, low impactHuman approvesWrite, high impactIMPACT IF WRONG →ACCESS TYPE →
Where to start: read-only access on low-impact questions.

Is MCP the same as an API, a plugin or an AI agent?

No. MCP is related to all three, but it is a different thing from each of them.

  • An API is the interface a tool offers to software. An MCP server often calls an API behind the scenes. MCP adds a standard way for an AI application to discover and use what the server offers, so the assistant does not need custom code for each API.
  • A plugin or connector is a product feature. In many assistants, a connector is how you turn an MCP server on. MCP is the standard underneath, and the connector is the packaged experience.
  • An AI agent is a system that plans and carries out multi-step work. An agent can use MCP to reach tools, but MCP does not make anything an agent. It only supplies the connections.

MCP is not a language model and not a security product: the specification says it cannot enforce its security principles at the protocol level, which leaves the safeguards to the applications and the people who set them up.

What are the risks of using MCP?

The main risks are that a connected assistant can reach more than you intended, that a server you install can run code or misuse access, and that content the assistant reads can try to steer it. The specification is direct about this: MCP enables powerful capabilities through arbitrary data access and code execution paths, and implementors must address the trust and safety considerations that come with them.

The specification states three principles. Users must explicitly consent to and understand all data access and operations, and must keep control over what is shared and what actions are taken. Hosts must obtain explicit user consent before exposing user data to servers. And tools represent arbitrary code execution, so hosts must obtain explicit user consent before invoking any tool, and descriptions of tool behavior should be treated as untrusted unless they come from a trusted server. On human oversight, the page on tools says there should always be a human in the loop with the ability to deny tool invocations.

The official security guidance lists concrete attack patterns. A few matter most to non-developers:

  • Untrusted local servers. A local server runs on your computer with the same privileges as the client. The guidance warns about malicious startup commands and says clients should show the exact command before running it. Install servers only from sources you can verify.
  • Over-broad permissions. The guidance describes scope minimization: broad permissions granted up front enlarge the damage if a token leaks, and it recommends starting with a minimal set of low-risk read operations and raising access only when needed.
  • Token misuse. A server must not accept tokens that were not issued for it, and passing tokens through to other services is explicitly forbidden because it weakens controls and audit trails.

The specification also advises that clients log tool usage and show tool inputs to the user before calling the server.

What does a safe setup look like for a marketing team?

A safe setup follows the same logic as any access you would give a new freelancer: start small, separate it from everything else and review before anything changes. These rules apply to any MCP connection, whatever the data source.

  1. Read-only first. Connect with access that can view data but not change it. Move to anything more only after the read-only version has proven useful.
  2. Least privilege. Grant the narrowest access that answers your questions: one account rather than all accounts, one property rather than the whole organization.
  3. Separate credentials. Use a dedicated connection for the assistant, not your personal admin login, so you can revoke it without touching anything else.
  4. A human approves changes. Any change to ads, budgets, tracking or live pages is reviewed by a person before it happens. Check that your host actually asks, rather than assuming it does.
  5. Verify the server. Prefer a server published by the vendor of the product or by a maintainer you can identify, read its documentation and check what permissions it asks for.
  6. Log and revoke. Keep a record of what the assistant did, review access on a schedule and remove connections you no longer use.
  7. Check the numbers. Compare a sample of the assistant's answers with the source report before using them in a decision.

Need a safe way to put AI on your reporting?

We help you define what an assistant may read, what it may never change and who reviews the output.

Plan your reporting workflow

How do you get started with MCP as a marketer?

The simplest start is one assistant, one read-only connection and one recurring question. You do not need to build anything to learn what MCP is useful for.

  1. Pick the question. Choose a question you already answer by hand every week or month, such as "what changed in organic clicks since last period?" A concrete question tells you which data source to connect first.
  2. Check what your assistant supports. Look in your assistant's documentation for how it connects to MCP servers. Our guide to using MCP with Claude walks through that for one assistant, and our piece on ChatGPT and MCP covers the other.
  3. Find the server for your data source and read who publishes it, what it can do and what access it needs. For each product we say whether an official server exists or only community ones, with the date, in the dedicated guides.
  4. Connect read-only, with a dedicated account or credential, and test with a question whose answer you already know.
  5. Compare the answer with the source report. If it matches, add one more question. If it does not, find out why before going further.
  6. Write down the rules. Record what the connection is for, who owns it and who reviews the output. A short document is enough.

If you have a developer or technical partner, you can go one step further and build a server for a source that has none. That is a different project from using one, and our guide on how to build an MCP server explains the scope and what to expect. For a view of where this fits among the wider changes in the field, see our article on AI in digital marketing.

Frequently Asked Questions

What does MCP stand for?

MCP stands for Model Context Protocol. It is an open-source standard for connecting AI applications to external systems such as data sources, tools and workflows. Anthropic introduced it in November 2024, and the project is documented at modelcontextprotocol.io.

What is an MCP server?

An MCP server is a program that provides context and capabilities to an AI application through the MCP standard. It can offer tools the model can call, resources it can read and prompts a person can start. A server can run on your own computer or be hosted by a vendor.

Do I need to code to use MCP?

Not to use it. Many assistants let you connect servers through settings, and you then work in plain language. Building a new server does need development skills. Some setups also ask for a command or a configuration step, so read the documentation of the server you choose.

Is MCP safe to use with my marketing accounts?

It can be, if you set it up with care. The specification asks hosts to obtain consent before data is shared or a tool runs, and it recommends a human in the loop. In practice that means read-only access first, the narrowest permissions, a separate credential and review before any change.

Which assistants support MCP?

The official introduction says AI assistants such as Claude and ChatGPT, and development tools such as Visual Studio Code and Cursor, support it. Support and features differ by product and by plan, and they change often, so check the current documentation of the assistant you use.

What is the difference between MCP and an API?

An API is how one piece of software talks to a service. MCP is a standard way for an AI application to discover and use what a server offers. An MCP server often uses an API underneath, and MCP means the assistant does not need a custom integration for each one.

Next steps

Pick one question you answer by hand every week and connect one data source read-only. Search data is the gentlest first connection to try. To plan which of your recurring reporting and operations tasks could run as an automated process, see our work on business process automation, and when you are ready to talk it through, get in touch with the questions you want answered and the tools you use.

Sources

Newsletter

One short email when we publish

Notes on AI, image, video and brand work, from the people who do it. No spam.

Author