HeyMark
Content

Blog

What is MCP in AI, and what is it for?

12 min read

MCP (Model Context Protocol) is an open standard for connecting AI applications, such as Claude or ChatGPT, to external systems: your files, a database, a calendar or the tool you use to run your social media. With an MCP server connected, the assistant looks up real information and gets tasks done in that tool, instead of only drafting text you then copy and paste. Anthropic introduced it as open source on November 25, 2024, and since December 2025 it has been a Linux Foundation project.

What MCP is for

You ask your assistant which of your Instagram posts did best last month. With no connection, it can’t see your account: it gives you advice that would fit anyone, or asks you to paste in the numbers. If the tool that holds those metrics offers an MCP server and you connect it, the assistant looks them up itself, answers with your actual posts and, if you ask, leaves a draft on your calendar.

Before MCP, every connection between a model and a data source was built from scratch. That is how Anthropic framed it at launch: models were cut off from data, and every new source needed its own integration, which made connected systems hard to scale. MCP proposes one shared protocol. The official documentation compares it to a USB-C port: a standard way to plug AI applications into external systems. Whoever builds a server builds it once, and any compatible application can use it.

The examples in that documentation include an agent that reaches into your Google Calendar and Notion to act as a more personal assistant, and Claude Code generating an entire web app from a Figma design. That is why MCP comes up so often in conversations about agentic AI: it is one of the ways an agent uses tools.

Who created MCP and who governs it today

  • Anthropic created it. On November 25, 2024, it published the specification and SDKs, added support for local MCP servers to the Claude Desktop apps and opened a repository of prebuilt servers, including ones for Google Drive, Slack, GitHub, Git, Postgres and Puppeteer.
  • Since December 9, 2025, it has been a Linux Foundation project. That day the foundation announced the Agentic AI Foundation (AAIF), with three founding projects: MCP, contributed by Anthropic; goose, from Block; and AGENTS.md, from OpenAI. According to that announcement, there were more than 10,000 published MCP servers at the time, and Claude, Cursor, Microsoft Copilot, Gemini, VS Code and ChatGPT, among other platforms, had adopted it.
  • Its maintainers set the technical direction. According to its governance page, those roles belong to individuals: no seats are reserved for any company. Changes to the specification are proposed in writing, and the code and specification are published under the Apache 2.0 license.

Host, client and server: the parts of MCP

MCP splits the work among three participants, according to its architecture:

  • Host: the AI application you use, such as Claude Desktop, Claude Code or Visual Studio Code. It coordinates the connections.
  • Client: the component the host creates to talk to a server. There is one client per server, each with its own connection.
  • Server: the program that offers data and capabilities to the client.

The documentation’s example: Visual Studio Code acts as the host. When it connects to the Sentry MCP server, it creates a client for that connection; if it also connects to a local filesystem server, it creates a second client.

What an MCP server is and what it offers

An MCP server is the program that exposes a service’s capabilities so an AI application can use them. It can run in two places:

  • Local: on your own computer, talking to the application through the system’s standard input and output (stdio). The filesystem server that Claude Desktop launches is an example. It usually serves a single client.
  • Remote: on the provider’s platform, over the internet, using the transport called Streamable HTTP. Sentry’s official server is an example. It usually serves many clients. To identify you, it accepts keys, tokens or custom headers, and the protocol recommends OAuth: you sign in to the service and approve the access.

What a server offers is organized into three parts, described in the server concepts:

Part What it is Who decides to use it
Tools Functions that do something: search flights, send a message, create an event. They can write to a database, call other APIs or modify files The model, based on what you ask
Resources Read-only data that serves as context: the contents of a file, a database schema, an API’s documentation The application
Prompts Ready-made instruction templates the server provides: plan a trip, summarize your meetings, draft an email You, when you pick one; many applications show them as commands that start with /

It also works the other way: through the application, the server can ask you for a missing detail or for confirmation before it carries out an action.

This is how a server describes one of its tools when the client asks what it has (the request is called tools/list). It is the example from the official documentation, shortened:

{
  "name": "weather_current",
  "title": "Weather Information",
  "description": "Get current weather information for any location worldwide",
  "inputSchema": {
    "type": "object",
    "properties": {
      "location": { "type": "string", "description": "City name, address, or coordinates" }
    },
    "required": ["location"]
  }
}

The model reads the description and, when you ask about the weather in a city, decides to use that tool. The application makes the call (tools/call), gets the server’s response and hands it to the model so it can answer you. That is why a clear description for each tool matters: it is what the model uses to choose.

MCP, API or plugin: what changes if you don’t code

An API is the door a program uses to talk to a service. To use it, someone writes code: what to request, with which permissions, what to do if something fails. An MCP server often calls an API underneath, but presents it in a way an assistant understands: which tools exist, what each one is for and what data it needs. You don’t build the integration. You add the server in your AI application, sign in and ask for tasks in your own words.

A plugin or extension is usually made for one specific application. An MCP server is built once and works with any compatible application, even if each one gives it a different name in its interface: in Claude, a remote server is added as a custom connector.

If you’re deciding how to connect your social media accounts to software, we compare the options in Social media MCP vs API: which one you need.

What you authorize when you connect an MCP server

Connecting a server gives your assistant access to whatever that server can read and do with your account. The specification puts control on your side: you must explicitly consent to access to your data and to actions, and the application must ask your permission before using any tool. It also makes clear that the protocol can’t enforce those principles on its own: it depends on the application and the server implementing them well. Before you connect one, check five things:

  1. Who is behind it. Claude’s help center warns that a custom connector links you to services Anthropic has not verified, and OpenAI’s documentation says the same about custom MCP servers in ChatGPT. The specification adds that tool descriptions are untrusted unless they come from a trusted server. Connect servers from organizations you trust, preferably the service’s official one.
  2. What permissions it asks for. On the authorization screen, look at what it requests and whether that matches what you want it to do. Reading your metrics and posting on your behalf carry different levels of risk.
  3. What it can do without asking you. Review every approval request before accepting it. Keep the always-allow option for tools you trust to work unsupervised, and turn off the ones you won’t use.
  4. Hidden instructions. A malicious server, or content the assistant reads through it, can carry hidden instructions to make it do something you didn’t ask for, including sending private data from one server to another. This is called prompt injection. Pay attention to what goes into and comes out of each tool.
  5. If it’s local, what command it runs. A local server runs on your computer with the same privileges as the application. The protocol’s security best practices require clients that offer one-click local server setup to show you the full command before running it: read it, and don’t accept it if you don’t know what it does. If you set it up by hand, you’re the one pasting the command: check it the same way.

When you stop using a server, revoke its access from the application.

An example: HeyMark’s MCP server in Claude

HeyMark is social media management software with an AI agent, and it runs a remote MCP server at https://mcp.heymark.ai. The URL is the same for everyone: what protects your account is the authorization you approve. In Claude it is added as a custom connector: you paste the URL, HeyMark opens in your browser, you sign in with your account and approve the access. From then on, Claude sees what your account sees in your brands and their connected networks (Instagram, TikTok, Facebook, LinkedIn and YouTube), with your permissions, and nothing else.

In terms of the parts above: Claude is the host, and adding the connector creates the client that keeps the connection to HeyMark, a remote server that uses Streamable HTTP and OAuth 2.1. Its tools are split into read and write, and some writes wait for your approval:

What you ask Claude Tool What happens
What do I need to reply to today? get_inbox_overview Read: how many unread messages there are and which conversations are waiting for a reply
How did I do over the last 30 days? get_analytics_overview Read: follower trend, totals against the previous period and top posts
Leave a draft for Thursday create_post Direct write: creates planning content. Scheduling and publishing are separate actions
Schedule it for Thursday at 10:00 schedule_post Direct write, with no preview, and it can publish later: check the date and content before asking for it
Publish it now, or reply to this comment publish_post, reply_to_comment Asks first: waits for you to approve a preview

The tools available depend on your permissions and your client. If you disconnect the connector, only the access goes away: scheduled posts and enabled automations keep running in HeyMark, and you stop them there if you want to. HeyMark has no public REST API: programmatic access is through MCP.

The full list of tools, which ones ask first and the technical details for client builders are on HeyMark’s MCP server page. The step-by-step setup is in HeyMark in Claude.

Plan your next post in HeyMark.

Keep the idea, review the draft with your team, and see how it performed in the accounts you connected.