Short answer
MCP (Model Context Protocol) is an open standard for connecting AI models to external tools and data, such as files, databases, GitHub, Slack or a web browser. Instead of building a custom integration for every app and every AI model, a developer writes one MCP server, and any MCP-compatible AI app can use it. Anthropic introduced MCP in November 2024; OpenAI, Google, Microsoft and many developer tools have since adopted it, which is why it is often called the "USB-C port for AI".
Why MCP exists
A large language model on its own only knows what was in its training data and can only produce text. To be useful at work, it needs to read your documents, query a database, create a calendar event or open a pull request. Before MCP, every AI app needed its own custom connector for every service. With M AI apps and N services, that meant building and maintaining up to M × N integrations.
MCP turns that into M + N: each service exposes one MCP server, each AI app implements one MCP client, and they can all talk to each other.
How MCP works
MCP uses a simple client-server design built on JSON-RPC 2.0 messages.
| Part | What it is | Example |
|---|---|---|
| Host | The AI application the user works in | Claude Desktop, ChatGPT, an IDE like Cursor or VS Code |
| Client | The connector inside the host that talks to one server | One client per connected server |
| Server | A small program that exposes capabilities of a tool or data source | GitHub server, file system server, Postgres server |
A typical flow:
1. The host starts and connects to the configured MCP servers. 2. Each server tells the client what it offers. 3. You ask a question, for example "Summarise the open bugs in our repo". 4. The model sees the available tools, decides to call one (such as list_issues) and the client sends the request to the server. 5. The server runs the action and returns the result, which is added to the model's context window. 6. The model writes its answer using the real data.
Servers can run locally (started by the host and connected over standard input/output) or remotely over HTTP, which lets companies host MCP servers for their products.
The three building blocks: tools, resources and prompts
- Tools are actions the model can call: search the web, run a SQL query, send a message, create a file. The model decides when to use them, usually with user approval.
- Resources are data the app can read and add to context: files, documents, database records, API responses.
- Prompts are reusable templates a server provides, such as "review this pull request" or "draft a release note".
This makes MCP the plumbing behind many AI agents: the agent plans, and MCP gives it hands.
What people use MCP for
- Coding: letting an AI assistant read a repository, run tests, check CI and open pull requests.
- Company knowledge: searching Google Drive, Notion, Confluence or SharePoint and answering with sources, a live alternative to building your own RAG pipeline.
- Data analysis: querying databases and spreadsheets directly instead of copying data into a chat.
- Productivity: reading email and calendars, creating tasks, posting updates.
- Browser and desktop automation: controlling a browser to fill forms or test a website.
- Design and content: connecting AI to design tools, CMSs and analytics.
MCP vs function calling vs plugins
| Function calling | MCP | |
|---|---|---|
| What it is | A model feature: the model outputs a structured tool call | A protocol: a standard way to discover and run tools |
| Where tools live | Defined in each app's code | In reusable servers any MCP host can connect to |
| Portability | Tied to one app and often one model API | Works across apps and models that support MCP |
| Discovery | Developer hard-codes the tool list | Hosts ask servers what they offer at runtime |
MCP does not replace function calling; it builds on it. The model still uses function calling to request a tool, and MCP standardises how that request reaches the tool.
Who supports MCP?
Anthropic open-sourced MCP in November 2024 with SDKs and reference servers. During 2025, OpenAI added MCP support to its Agents SDK and ChatGPT, Google announced support for Gemini, and Microsoft built MCP into GitHub Copilot, VS Code and Windows. In December 2025, Anthropic donated MCP to a new Linux Foundation project, the Agentic AI Foundation, so it is governed as a neutral open standard. Thousands of community and official servers now exist for popular services.
Security risks to know
Connecting AI to real systems brings real risks:
- Prompt injection: text inside a web page, email or document can contain hidden instructions that try to make the model misuse its tools.
- Over-broad permissions: a server with write access to everything can do a lot of damage if the model is tricked or makes a mistake.
- Untrusted servers: a malicious MCP server can describe its tools misleadingly or leak data.
- Data exposure: results returned by tools go into the model's context and may be logged.
Good practice: install servers only from trusted sources, grant the least access needed, require confirmation for actions such as sending, deleting or paying, and review what each tool does. This is part of the wider AI alignment and safety challenge of giving models more autonomy.
How to start using MCP
1. As a user: open your AI app's settings (for example Claude Desktop, ChatGPT or your code editor), find the connectors or MCP section, and add a server such as file system, GitHub or a database. 2. As a developer: use an official SDK (Python, TypeScript and others) to write a server that exposes a few tools, then connect it to a host for testing.
A minimal Python server looks like this:
from mcp.server.fastmcp import FastMCP
mcp = FastMCP("gpu-prices")
@mcp.tool()
def h100_price() -> str:
"""Return today's typical H100 cloud price per GPU-hour."""
return "See https://tensorcore.si/gpu-cloud-pricing/ for live prices"
if __name__ == "__main__":
mcp.run()Frequently asked questions
Is MCP only for Claude?
No. Anthropic created it, but MCP is an open standard supported by many AI apps and models, including products from OpenAI, Google and Microsoft, and open-source tools.
Do I need to code to use MCP?
No. Many apps let you add ready-made MCP servers or connectors from a settings menu. Coding is only needed to build your own server.
Is MCP the same as an API?
MCP usually sits on top of existing APIs. An MCP server wraps a service's API and describes it in a standard way that AI models can discover and use.
Does MCP make AI hallucinate less?
It can, because the model can fetch real data instead of guessing. But the model can still misread results, so important outputs should be checked. See AI hallucinations.
Sources and further reading
- Anthropic, Introducing the Model Context Protocol (November 2024)
- Anthropic, Donating the Model Context Protocol and establishing the Agentic AI Foundation (December 2025)
- Model Context Protocol documentation and specification
- Model Context Protocol on GitHub




Leave a Reply