What is MCP?

MCP — the Model Context Protocol — is the standard that lets an AI assistant use your other software. It is why you can now ask Claude or ChatGPT to do something in a tool you already pay for, rather than copying text back and forth between tabs.

If you work in marketing rather than engineering, almost everything written about MCP is aimed at the wrong audience. Here is the version that matters to you.

The problem MCP solves

Until recently, an AI assistant was a very capable thing in a box. It could write you a LinkedIn post, and then you copied that post, opened another tab, pasted it, and scheduled it. Everything the assistant did ended at the edge of the chat window.

The obvious fix — let the assistant use other software — ran into a combinatorial problem. Every assistant would need a custom integration with every tool. Five assistants and a thousand tools is five thousand integrations, none of which anyone would build.

MCP is the agreement that stops that. It defines one way for an assistant to ask “what can you do?” and one way to call the answer. A tool implements it once and works with every assistant that speaks it. An assistant implements it once and can use every tool.

It is the same idea as USB. Nobody celebrates USB, and that is rather the point: standards are boring when they work.

What an MCP server actually is

An MCP server is a service that publishes a list of things an assistant can ask it to do. Each thing is called a tool, which is slightly confusing because the whole product is also colloquially a tool. In MCP terms, a tool is one specific action.

A LinkedIn publishing server might expose:

• get_context — tell me about this user’s brand and voice
• create_post — save this text as a draft
• schedule_post — queue this for Tuesday at 9am
• publish_post_now — put this on LinkedIn immediately

When you ask your assistant to “write something about our new pricing and schedule it for Tuesday”, it reads that list, works out that it needs the last one, writes the post, and calls schedule_post with the text and the time.

You can see a real, complete example on our own tool reference page — nine tools, what each does, and what each costs.

Connector, app, server: same thing

The naming is a mess and it is worth clearing up, because it makes the documentation confusing.

• Claude calls them connectors.
• ChatGPT calls them apps or connectors depending on where you are looking.
• The standard calls them MCP servers.

These are the same thing described from different angles. If a product tells you it has “a connector for Claude”, it is running an MCP server. If a guide tells you to “add a custom MCP server”, you are looking for the connectors screen.

There is one real distinction underneath the naming: hosted versus self-hosted. A hosted server is run by the company whose product it is — you paste a web address and sign in, and there is nothing to install. A self-hosted one runs on your own machine, which means installing and maintaining it, and means it is unreachable when your laptop is shut. Most commercial products are hosted. Most of the open-source MCP servers you will find on GitHub are not.

What an MCP connector can and cannot see

This is the question worth being precise about, because the answer is reassuring and most people assume worse.

An assistant with a connector does not get access to your account. It gets access to the specific tools that server exposes, and it receives only what those tools return.

Concretely, with a LinkedIn publishing connector:

• It can call schedule_post and get back “scheduled for Tuesday 9am”.
• It cannot read your LinkedIn inbox, because there is no tool for that.
• It cannot see your other drafts, your analytics or anyone else’s data, because those are not what the tools return.
• It never holds a LinkedIn credential at all. Your LinkedIn authorisation lives with the publishing tool; the assistant only ever authenticates to that tool.

And you can narrow it further. Every serious MCP client lets you set permissions per tool, so you can allow drafting and forbid publishing, or require approval on anything that changes something. That control is the part worth actually using rather than clicking past.

Where the real risk is

The protocol does not make anything safe by itself. It is a way of talking, and it will carry a bad idea just as faithfully as a good one.

The distinction that matters, in this category specifically:

Servers that authorise properly. You sign in through the service’s own login, it issues a token scoped to specific actions, and the server can do those things and nothing else. Revoking it takes one click.

Servers that want your session cookie. Some LinkedIn MCP servers — including several of the most-starred open-source ones — ask you to paste the li_at cookie from your browser. That is your entire logged-in session: messages, connections, everything, with no scoping and no per-action limit. It also drives your account in a way LinkedIn detects and restricts accounts for.

If a setup guide asks for a cookie rather than sending you to a login page, that is the whole signal. It is not a technicality; it is the difference between granting permission and handing over the keys.

Why marketers should care

Two things change, and only one of them is obvious.

The obvious one: less copying between tabs. Write and publish in one place. Real, and modest.

The one that matters more: the context is already there. The reason this is better than a scheduling app is not the keystrokes saved. It is that in a conversation where you have just worked through a problem — a pricing decision, a customer call, a thing that went wrong — the material for the post already exists in that chat. “Write that up” is a far better prompt than opening an empty composer three hours later and trying to reconstruct what was interesting.

That is a genuinely different workflow, and it is why MCP is more interesting for content than the feature list suggests.

The catch nobody mentions

If the assistant writes the post, the post sounds like the assistant. There is now enough machine-written text on LinkedIn that readers clock it quickly, and a post that reads as generated earns less than one that does not.

So the useful shape is: the assistant decides what to say, and something with a model of your voice decides how you say it. That is a design decision a good marketing MCP server makes for you — WordPush’s write_post hands drafting to a voice built from an interview you did at signup, rather than to the assistant. More on that in how to use AI to write LinkedIn posts.

Getting started

If you want to try it, the shortest path is Claude, which supports custom connectors on every plan including Free. Our Claude setup guide is about three minutes, and can Claude post to LinkedIn covers what to expect once it is connected.