What is MCP, the Model Context Protocol? How MCP servers work, MCP vs function calling vs APIs, and how to build one, from running a public server

What is MCP, the Model Context Protocol?

MCP is an open protocol, introduced by Anthropic in late 2024 and now maintained as an open standard, that defines how an AI application connects to external tools, data sources and prompt templates through a common interface [1][3]. A server exposes capabilities; a client inside the AI application discovers and uses them; the model decides when to call them. The point is reuse: a server for your database, your docs or your tool set is written once and works with every client that speaks MCP, instead of a custom integration per assistant.

Also asked as: what is mcp model context protocol · what is mcp in ai · mcp meaning ai · model context protocol explained · what does mcp stand for · mcp protocol explained · what is an mcp server · mcp ai explained simply · anthropic mcp

I run a public MCP server at tools.scult.in that exposes the same free tools and the prompt library the website offers, so an agent can call, say, the schema generator or fetch a stamped prompt with one line of client configuration and no signup [9]. Building it taught me what the protocol is for in practice: the client side is finished by other people, so a server is the whole product. This page is what I wish I had read first, and the longer build note is on the portfolio [10].

Before MCP, every tool was an integration project per assistant. After MCP, a tool is a server, and the assistants come to it. Pranjul Rathour, from running the tools.scult.in MCP server

How does MCP work?

Three parts. The host is the AI application, a coding assistant, a chat app, an agent runtime. Inside it, an MCP client connects to one or more MCP servers, over standard input and output for local servers or HTTP for remote ones, using JSON-RPC messages [6]. The server declares what it offers: tools with JSON schemas, resources with URIs, prompt templates. The client lists them and hands the tool definitions to the model. When the model emits a tool call, the client forwards it to the server, the server executes and returns a result, and the result goes back into the model's context.

Also asked as: how does mcp work · mcp architecture · mcp client server · mcp host client server · how mcp servers work · mcp transport stdio http · mcp json rpc · mcp lifecycle

What are tools, resources and prompts in MCP?

Tools are functions the model can call, each with a name, a description and a JSON schema for arguments; they are for actions and computations. Resources are data the client can read by URI, files, database rows, documents, for context rather than action. Prompts are reusable templates the server offers, with arguments, so a client can present them as slash commands or presets. Most servers are tools-first; resources and prompts are where the protocol gets interesting for knowledge and workflow servers.

Also asked as: mcp tools resources prompts · what is a tool in mcp · mcp resources explained · mcp prompts · mcp tool schema · mcp primitives · mcp capabilities

What is the difference between MCP and function calling?

Function calling is a feature of a model provider's API: you send tool definitions, the model returns structured tool calls, you execute them [4][5]. MCP is a protocol between the application and tool servers. They compose rather than compete: the MCP client fetches tool definitions from servers and passes them to the model through function calling; when the model calls one, the client routes the call to the right server. Without MCP you write and host every tool inside your app. With MCP, tools live in servers anyone can plug in.

Also asked as: mcp vs function calling · mcp vs tool calling · is mcp the same as function calling · function calling vs mcp server · mcp vs openai functions · mcp vs langchain tools · mcp vs plugins

MCP vs API: do I still need a REST API?

Yes, usually underneath. An MCP server typically wraps an existing API or database and exposes a handful of model-friendly tools with descriptions that say when to use them. The REST API stays for your own front end and for programmatic clients; the MCP server is the model-facing surface. The design skill is in the wrapping: three well-described tools that map to user intents beat thirty tools that mirror your endpoints one to one.

Also asked as: mcp vs api · mcp vs rest api · do i need mcp · is mcp replacing apis · mcp server vs api wrapper · mcp for existing api · openapi to mcp

How do I build an MCP server?

Pick the official SDK in your language, define two or three tools with tight descriptions and JSON schemas, implement each as a function that returns text or structured content, and run the server over standard input and output for local use or HTTP for remote use [2]. Test it with a client you already use, a coding assistant that supports MCP, before writing more tools. Add resources and prompts only if a client would use them. Publish with a README that shows the one-line client configuration.

Also asked as: how to build an mcp server · mcp server tutorial · create mcp server python · mcp server typescript · mcp server example · mcp sdk · how to make mcp server · mcp server development · fastmcp

The pattern for a tool definition, in the shape the SDKs use:

{
  "name": "get_prompt",
  "description": "Fetch a version-stamped prompt from the tools.scult.in library by slug. Use when the user asks for a tested prompt for a task such as a RAG system prompt or a code review checklist.",
  "inputSchema": {
    "type": "object",
    "properties": {
      "slug": {"type": "string", "description": "Prompt identifier, e.g. rag-system-prompt-v3"}
    },
    "required": ["slug"]
  }
}

The description is the interface. Models pick tools by reading it, and a vague description is the most common reason a good tool never gets called.

How do I use MCP servers in a coding assistant or chat app?

Add the server to the client's configuration, usually a JSON file or a settings page, with the command to start it or the URL to reach it. The client connects, lists the tools, and they appear to the model in every conversation. Start with one server; too many tools dilute the model's choices. Prefer servers with a few sharp tools and clear descriptions. Remove servers you stop using.

Also asked as: how to use mcp servers · mcp server configuration · add mcp server to claude · mcp server setup · best mcp servers · mcp servers list · mcp server for coding · mcp client

What are the security risks of MCP?

The same as for any agent with tools, concentrated. A tool with write access can be triggered by a prompt injection hidden in a document the model read. A malicious or compromised server can return content that steers the model. Servers run with whatever permissions you give the process. Defences: least-privilege tools, read-only by default; explicit confirmation for anything destructive or outward-facing; treating every tool result as untrusted data; pinning and reviewing the servers you install; logging every call. OWASP's LLM risk list applies directly [7].

Also asked as: mcp security · is mcp secure · mcp server security risks · mcp prompt injection · mcp permissions · mcp tool poisoning · secure mcp server · mcp authentication

When should I use MCP, and when not?

Use MCP when more than one AI client will use the same tools, when you want users to plug your product into their assistant, or when you want to reuse servers others have built. Skip it when a single application calls its own functions and nothing else ever will; plain function calling inside the app is simpler. Anthropic's guidance on agents applies here too: the simplest architecture that works, and tools only where the model's judgement is actually needed [8].

Also asked as: when to use mcp · mcp use cases · should i build an mcp server · mcp benefits · mcp advantages and disadvantages · mcp for enterprise · is mcp worth it

What does running a public MCP server teach?

That descriptions matter more than code, that clients differ in how they present tools, that logs are the only way to learn what models actually call, and that a server with no signup and no keys gets used in ways you did not plan. The tools.scult.in server exposes the free tools and the 1,211-prompt library; the most-called tool is not the one I expected, and the description edits I made from the logs improved call accuracy more than any code change [9][10].

Also asked as: public mcp server · mcp server examples · real world mcp server · mcp server lessons · open mcp server · mcp server without authentication

MCP interview questions

Explain the host, client and server roles. Compare MCP with function calling and say how they compose. Name the three primitives and an example of each. Describe the security model and the two biggest risks. Explain what makes a tool description good. Say when you would not use MCP. Having built even a two-tool server turns all of these into stories.

Also asked as: mcp interview questions · model context protocol interview · ai agent interview mcp · tool use interview questions

Where should I start?

Point an MCP-capable client at the server at tools.scult.in and use a tool in a real conversation [9]. Then build a two-tool server over something you own, a folder of notes, a small database, test it from the same client, and read the logs. For a hands-on session on agents, tool calling and MCP at your college or team, email pranjulrathour41@gmail.com or use pranjulrathour.scult.in/invite.

Pranjul Rathour

GenAI Engineer · Kanpur, Uttar Pradesh, 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 Kanpur) and BYTEBATTLE (MeetKats).

Open to GenAI roles, hackathon judging, mentorship sessions and guest talks at colleges. Email pranjulrathour41@gmail.com.

PortfolioInvite me to your campusHire meLinkedInGitHubXInstagramThreadsFacebookBlueskyDev.toBloggerHashnodeSCULTEmail