What Is an MCP Server? The Model Context Protocol Explained
How LLMs get tools – explained with the MCP server that runs this blog
This article is part of the AI Agents Guide – the curated learning path for AI agents.
A large language model is, at its core, a text generator: it predicts the next word. Impressive – but on its own it can't touch anything in the real world. It doesn't know your files, can't query a database and can't publish anything. This is exactly where the Model Context Protocol (MCP) comes in.
The problem: the LLM is locked in
As long as an LLM only sits in the chat window, it's a clever conversation partner without hands. But the moment it should become genuinely useful – change code in your repo, query a server's status, create a blog article – it needs a bridge to real tools. In the past every vendor built that bridge themselves, incompatibly. MCP standardizes it.
MCP: the USB standard for AI tools
The most fitting analogy: MCP is to AI tools what USB is to peripherals. Instead of building a separate adapter for every combination of LLM client and tool, both speak a shared protocol. An MCP server provides capabilities; an MCP client (e.g. Claude Desktop, Claude Code or LibreChat) uses them. Every server works with every client.
The flow is simple: the client asks the server “what tools do you have?”, the server answers with a list plus descriptions. The LLM then decides on its own when to call which tool – and the server executes it.
The three building blocks: tools, resources, prompts
- Tools are actions the LLM can trigger – functions with parameters (“create an article”, “query the click stats”). This is the most important and most common building block.
- Resources are data to read – files, database entries, API responses that the model receives as context.
- Prompts are predefined templates a client can offer as ready-made commands.
Concretely: my MCP server for devmaker.net
This isn't theory – this blog is run exactly this way. I built an MCP server that controls the Wagtail/Django website behind devmaker.net. Its tools are called, for example, write_article, set_streamfield_content, generate_ai_image or publish_page. An LLM sits in front of it as the client: I say “create the article and set the tags”, the model picks the right tools and calls them – the server does the actual work in the database.
A minimal server with FastMCP (Python) shows how little it takes:
from fastmcp import FastMCP
mcp = FastMCP("my-server")
@mcp.tool()
def server_time() -> str:
"""Returns the current server time as an ISO string."""
from datetime import datetime
return datetime.now().isoformat()
if __name__ == "__main__":
mcp.run()The docstring is not decoration: it's the description the LLM uses to decide whether to call the tool. With MCP, good descriptions are half the battle.
A server like this barely needs any resources – my devmaker MCP server runs as a lean service on the same power-efficient mini PC that carries the rest of my homelab:
Ad · Affiliate link – if you buy through it, I may earn a commission. It doesn’t change the price for you.
When is your own MCP server worth it?
Whenever you want an LLM to work with your own system repeatedly: your API, your database, your homelab, your tooling. For a one-off question, an MCP server is overkill. But if the model should perform concrete actions reliably and repeatably, a cleanly defined MCP server is the more robust route than any copy-paste into the chat.
What I left out
Deliberately left out for the overview: the transport topic (stdio vs. HTTP/SSE) and the pitfalls of remote operation – there's a dedicated article for that: MCP server doesn't survive a deploy: SSE vs. Streamable-HTTP. Likewise authentication and error handling, which a production server needs.
Conclusion & outlook
An MCP server is the standardized bridge that turns an LLM from a pure text machine into an actor that operates real tools. The concept is manageable – tools, resources, prompts – and with FastMCP you start in a few lines. If you now want to build one yourself, it continues step by step here: Developing MCP servers in Python: from empty repo to the first tool.
Ad · Affiliate link – if you buy through it, I may earn a commission. It doesn’t change the price for you.