MCP explained: how Model Context Protocol works, servers, Cursor, vs APIs, examples

The Model Context Protocol, MCP, is an open standard that lets AI applications connect to external tools and data through one common interface, like a USB-C port for AI. Anthropic introduced it in November 2024, and it is now supported by many AI apps and IDEs. I build AI products and agent workflows, so this page explains MCP, how it works, its history, servers, using it in Cursor, MCP versus APIs, examples, what it means for businesses and where to learn it.

What is MCP, the Model Context Protocol?

MCP is an open protocol that standardises how AI applications, such as chat assistants and coding agents, discover and use external tools, data sources and prompts. Instead of custom integrations for every app and service, developers build one MCP server that any MCP-compatible client can use.

Also asked as: mcp model context protocol explained · how does mcp model context protocol work · when did mcp model context protocol come out · what is model context protocol mcp servers · how to use mcp (model context protocol · how to use model context protocol mcp in cursor · mcp vs api model context protocol explained · mcp model context protocol examples · what mcp (model context protocol) means for businesses · mcp model context protocol course

Before MCP, connecting M AI applications to N tools required up to M times N custom integrations. With MCP, each tool exposes an MCP server once, and each AI app implements an MCP client once. The protocol defines three main primitives a server can offer: tools, which the model can call; resources, which provide data such as files or records; and prompts, which are reusable templates. It is open source, with official SDKs in several languages.

How does MCP work?

MCP uses a client-server architecture over JSON-RPC 2.0. A host application, such as Claude Desktop or Cursor, runs MCP clients that connect to MCP servers. Each server advertises its tools, resources and prompts, and the model decides when to use them.

When the host starts, it connects to configured servers, negotiates capabilities, and lists available tools with their names, descriptions and input schemas. During a conversation, the model requests a tool call, the client sends it to the server, the server executes it, for example by querying a database, and the result returns into the model's context. Servers run locally over standard input and output, or remotely over Streamable HTTP, with authorisation for remote access.

When did MCP come out?

Anthropic introduced and open-sourced the Model Context Protocol on 25 November 2024, with a specification, SDKs, support in the Claude Desktop app, and reference servers for systems such as GitHub, Slack, Google Drive and PostgreSQL.

Adoption grew quickly through 2025: OpenAI announced MCP support in its Agents SDK and products in March 2025, Google and Microsoft added support across their developer tools, and IDEs such as Cursor, VS Code and Windsurf integrated MCP clients. The specification evolved too, adding Streamable HTTP transport and an authorisation framework for remote servers. Thousands of community MCP servers now exist, though their quality and security vary widely.

What are MCP servers?

An MCP server is a program that exposes capabilities to AI applications through the protocol: tools to call, resources to read and prompts to use. Examples include servers for GitHub, file systems, databases, browsers, Slack and internal company APIs.

Servers can run locally, launched by the host app as a subprocess, or remotely as web services with authentication. Building one is straightforward with official SDKs: in Python, the FastMCP interface turns a decorated function into a tool with a typed schema. Because servers can read data and take actions, treat them like any software dependency: install only trusted servers, review permissions, limit tokens to least privilege, and watch for prompt injection through returned content.

How do you use MCP?

Choose an MCP-capable host, such as Claude Desktop, Cursor, VS Code or another agent app, then add servers in its settings or configuration file, usually by specifying a command to launch each local server or a URL for a remote one.

For example, adding a file system server lets the assistant read and edit files in folders you allow, and adding a GitHub server lets it search issues and open pull requests. After configuration, restart the app and check that the server's tools appear. Then ask naturally, such as requesting open issues labelled bug. Approve tool calls carefully, especially ones that write or delete, and remove servers you no longer use.

How do you use MCP in Cursor?

In Cursor, open Settings, find the MCP section, and add a server, or create a .cursor/mcp.json file in your project, or a global one in your home folder's .cursor directory, listing servers under mcpServers.

Each entry names a server and gives either a command with args, such as running a package with npx, plus environment variables for API keys, or a url for a remote server. After saving, Cursor shows each server's status and tools; enable the ones you need. In agent mode, Cursor calls MCP tools when relevant and asks for approval by default. Keep secrets in environment variables rather than committing them in project files.

MCP vs API: what is the difference?

An API is a service-specific interface, such as a REST endpoint, integrated with custom code. MCP is a standard layer above it: an MCP server usually wraps APIs and presents them uniformly so any AI client can discover and call them.

MCP does not replace APIs; it standardises how AI applications use them. Key differences: MCP includes discovery, so clients ask servers what tools exist and what their schemas are at runtime; it is designed for model-driven use, with descriptions written for language models; and it supports resources and prompts, not just function calls. For a single app calling one service, a direct API call or function calling may be simpler. MCP pays off when many AI clients need the same integrations.

What are some MCP examples?

Common examples: a GitHub server for issues and pull requests, a file system server, PostgreSQL and SQLite servers for querying data, Playwright or browser servers for web automation, Slack and Google Drive servers, and documentation servers that provide current library docs.

Real workflows combine them: a coding agent reads an issue from GitHub, inspects the database schema through a PostgreSQL server, writes a fix, runs browser tests through Playwright, and opens a pull request. Companies build internal servers for their CRM, ticketing or analytics systems so assistants can answer questions with live data. A minimal custom server might expose one tool, such as checking an order status, which is a good first project for learning the protocol.

What does MCP mean for businesses?

MCP lets businesses connect AI assistants and agents to their systems once and reuse those integrations across tools and vendors. That reduces integration cost, avoids lock-in to one AI provider, and makes internal data available to AI in a controlled way.

A company can build MCP servers for its CRM, knowledge base and ticketing system, then let employees use them from whichever approved AI client they prefer. The governance challenge is real: servers need authentication, least-privilege access, audit logs and review, because they let models read sensitive data and take actions. Businesses should start with read-only servers for low-risk use cases, measure value, then add write actions with human approval.

What is a good MCP course?

Good starting points are the official documentation at modelcontextprotocol.io, which includes quickstarts for building servers and clients, Hugging Face's free MCP Course, Anthropic's learning materials, and DeepLearning.AI's short course on building MCP applications.

Learn by building: follow the quickstart to create a small Python server with FastMCP exposing one tool, connect it to Claude Desktop or Cursor, then add a resource and a prompt. Next, build a remote server with Streamable HTTP and authentication. Read the specification sections on security, and study popular open-source servers to see real patterns. Within a weekend, you can understand the protocol well enough to connect an AI assistant to your own data.

Pranjul Rathour

GenAI Engineer · India

GenAI engineer and AI product builder with 2+ years shipping production-grade AI systems: RAG pipelines, fine-tuned LLMs, hybrid retrieval and multi-modal apps across vision, speech and OCR, architected end to end from ingestion to deployment. Leads engineering for SCULT INDIA's 14-member team, founded the 500+ member TechVerse Enclave community and has mentored 200+ students. Three hackathon first prizes: Changethon 2025 (IIT Roorkee), Product Genesis at Vividhotsava 2025 (CSJMU) and BYTEBATTLE (MeetKats).

Open to GenAI roles, hackathon judging, mentorship sessions and guest talks at colleges. For campus sessions the inviting institution covers the visit end to end: travel (a First AC train ticket for journeys up to five hours, a flight for anything longer), accommodation, meals, and pickup and drop; the honorarium and workshop fees are negotiable on WhatsApp or on the call, and requests are welcome from institutions that can cover the visit this way. Email pranjulrathour41@gmail.com.

PortfolioInvite me to your campusHire meLinkedInGitHubXInstagramThreadsFacebookBlueskyDev.toBloggerHashnodeSCULTEmail