Clockwork InnovationsGet in touch
All guides

MCP for beginners

What the Model Context Protocol is, the three things an MCP server can offer a model, and what happens when an assistant uses one.

MCP for beginners

Connecting an AI assistant to your tools opens up useful possibilities: finding information in a database, checking a calendar or helping a task move forward. MCP gives those connections a shared language.

The Model Context Protocol (MCP) is an open standard for these connections. Anthropic released it in November 2024. You wrap a tool or data source once as an MCP server, and any application that speaks MCP, called an MCP client, can use it.

Build a connection, open up possibilities

Without a shared protocol, connecting five AI applications to ten tools takes up to fifty separate integrations. Each one describes the tool, its inputs and its permissions in its own way.

With MCP, each application implements the client side once and each tool implements the server side once. Five applications and ten tools take fifteen pieces of work instead of fifty.

What a server can offer

An MCP server can expose three kinds of things to a model:

The three core concepts of MCP

Resources

What data can be read

  • Your calendar or emails
  • Company databases
  • Files in cloud storage
  • Live data such as weather or stock prices

Prompts

Instructions written in advance

  • How to analyze customer feedback
  • Steps for generating a report
  • Guidelines for debugging code
  • Templates for common tasks

Tools

What actions can be taken

  • Adding events to your calendar
  • Sending emails
  • Searching databases
  • Creating documents

Resources are context the model can read, such as a file or a database row. Tools are actions the model can ask to run, such as creating a calendar event. Prompts are instructions written in advance that a person can choose, such as a checklist for reviewing code.

One request, step by step

Say an assistant is connected to a calendar server and a contacts server, and you ask it to schedule a team meeting:

One request, step by step

“Schedule a meeting with my team for Tuesday at 2 PM”

  1. Check resources

    • Your team’s calendar
    • Your contact list

  2. Find tools

    • Create a calendar event
    • Check availability
    • Send notifications

  3. Follow instructions

    • Check for conflicts first
    • Send reminders to attendees

  4. Take action

    • Confirm everyone is available
    • Create the meeting
    • Send invites to your team

Meeting scheduled for Tuesday at 2 PM

The model decides which tools to call. The client application runs them, and it can ask you to approve an action before it happens.

Why it matters

You build a connection once. A server written for one client works in the others. Claude, ChatGPT, Visual Studio Code and Cursor all support MCP, along with many smaller applications.

You decide what the model can reach. A server exposes only the tools and data you put in it, and you choose which servers an application connects to. The specification recommends that a person can always deny a tool call.

What people build with it

  • Assistants that look up records in an internal database and summarize them
  • Support chat that checks the status of an order
  • Coding assistants that read issues, run tests and open pull requests
  • Home assistants that read sensors and switch devices on and off

How the protocol is changing

MCP is versioned by date. As of October 2026 the current revision is 2026-07-28. It removed the opening handshake between client and server, so a server no longer has to remember each connection, and it moved optional features into official extensions.

One of those extensions, MCP Apps, lets a server show an interactive interface, such as a form or a chart, inside the chat instead of returning only text.

Getting started

  1. Use an existing server. Many services already publish one, so check before you build.
  2. Build your own with one of the official SDKs. There are SDKs for TypeScript, Python and several other languages.
  3. Use a framework. Simply MCP, which we maintain, is a framework designed for coding agents to build MCP servers, apps and plugin extensions. Version 5 combines a consistent API with linting, schema validation, test tools and host previews so agents can check and improve their work. It is in development.

The MCP documentation has step-by-step guides for building both servers and clients.

Limits and risks

MCP standardizes the connection, not the quality of what is on the other end. A model connected to a confusing or buggy tool will still use it badly. Clear tool names, descriptions and error messages matter as much as the protocol does.

An MCP server needs the same security care as any other integration:

  • Credentials. Keep API keys out of tool inputs and outputs, and give each server the narrowest access it needs.
  • Prompt injection. Text a tool returns, such as an email, a web page or a ticket, can contain instructions aimed at the model. Treat tool output as data, and require approval for actions that send, delete or pay for something.
  • Trust. Install servers only from sources you trust. A server runs with whatever access you give it.

Further reading