What Is an MCP Server? The Magic Problem, Explained.

In 1973, famed scientist and fiction writer Arthur C. Clarke penned the phrase

Any sufficiently advanced technology is indistinguishable from magic.

At the time it was a tongue-and-cheek observation about human hubris. Today, it reads like a spec sheet for LLMs.

You ask Claude a question about your pipeline. Thirty seconds later, having asked Attention, it tells you which deals are at risk this week, the sum of their ARR, which reps are getting hit hardest on pricing objections, and where your team should focus Monday morning. All with citations from specific calls.

You did not write a query. You did not open a dashboard. You typed a sentence. It feels, as Arthur Clarke said it would… like magic.

But there is a problem with employing “magic.” When you cannot see what is happening inside the proverbial hat, you cannot tell a good trick from a bad one. You cannot verbalize why one AI experience feels sharp and another feels like talking to a confused intern. You cannot explain why the same prompt burns a negligible number of credits with one vendor’s connector and eats half your daily limit with another. You cannot evaluate whether a vendor’s “connect to Claude” announcement is a serious engineering investment or a logo slapped on a thin wrapper.

The thing inside the hat is called an MCP server. Once you understand what it actually is and how it actually works, the magic does not disappear. It just becomes something you can reason about, buy wisely, and hold vendors accountable for.

This article is the first of three. It aims to reveal the magic to the everyday audience. The next piece walks through how MCP servers fail and why tool design matters more than tool count. The third is a practical checklist for evaluating any MCP server before you connect it.

To understand the MCP server, we are going to shift from the mystique of theatre magic to one of the world’s most successful and standardized operational systems: the McDonald’s kitchen. The difference between an AI model that can pontificate on a complex task and one that can actually perform it is the difference between a brilliant food philosopher and a line cook operating in a fully standardized McDonald’s environment.

The Analogy in One Glance

• The customer is the user — the person in the dining room ordering a meal.
• The order ticket is the prompt — the natural-language request that travels from the dining room into the kitchen.
• The line cook is the AI model — a capable generalist who arrives already knowing the standard, but not the kitchen.
• The kitchen is the MCP server — a specific vendor’s setup, mapped out precisely according to the universal protocol. McDonald’s has one kitchen, Burger King has another, KFC has a third. There are thousands of kitchens because there are thousands of brands wanting to be reachable by AI.
• The equipment in the kitchen — the clam-shell grill, the bun toaster, the sauce dispenser, etc. is the underlying API. These are the actual capabilities the vendor exposes for the cook to use.
• The MCP protocol is the universal operating standard that defines what “compliant” means. It says: write your instructions in this format. Label your equipment using these conventions. Lay out your stations in this way. It is not McDonald’s specific. It is the rulebook every brand follows if they want a trained cook to be able to walk in and produce real work.
• The finished tray going back to the customer is the response — the synthesized answer Claude cooks up.

TL;DR

• MCP stands for Model Context Protocol. It is an open standard, published by Anthropic in November 2024 and now stewarded by the Agentic AI Foundation, a directed fund under the Linux Foundation. It defines a universal language for AI models to communicate with external software.
• An MCP server is the brand-specific training documentation and centralized operational system a company uses for efficient employee onboarding to their unique tools and products. There is one protocol and thousands of servers. Anthropic’s own Claude connector directory lists 75+ connectors, with over 10,000 public MCP servers active across the broader ecosystem.
• Without an MCP server, an AI model is like a brilliant person who can pontificate on the history and architecture of a Big Mac but cannot actually build one. The server provides the tools and the specific knowledge to actually build the burger.
• Not all MCP servers are built equally. A well-implemented one allows the AI to onboard instantly, understand every specific tool in the kitchen, and know exactly when to use a clam-shell grill versus a standard flat-top.
• System design directly determines your costs. A well-designed MCP server takes the burden off the AI’s context budget, letting it focus on the final assembly rather than hunting for where the pickles are kept.

The Big Mac Philosopher

A large language model is, at its core, a very capable but very isolated thing. It can write, reason, summarize, translate, and produce code at a level that impresses even the people who built it. What it cannot do, on its own, is open your inbox, check your calendar, query your data warehouse, look up a deal in Salesforce, or pull the transcript of last Tuesday’s call with your biggest prospect.

It is a brilliant person standing in an empty room. It is, in our analogy, a brilliant gourmand who could opine for hours on how a Big Mac is constructed… while having no ability to serve you lunch.

The problem with APIs, when it comes to AI models, is that APIs were designed for deterministic callers. Another piece of software that has read the documentation, written integration code, and knows exactly which endpoint to call with exactly which parameters. APIs assume the caller has done its homework in advance, like a line cook who has worked at a given McDonald’s for years.

An AI model does not do its homework in advance. It receives a question in plain English and has to figure out, at runtime, which tool to reach for and how to use it. It is closer to a new hire walking into a restaurant for the first time — capable, but needing to look around and find the standardized stations before they can produce anything. Critically, they have to do this “look around” and read the manual each and every time you ask them to cook an item.

That is a fundamentally different kind of caller, and it is the reason a standardized onboarding system was needed.

The Onboarding System and the Kitchen

In November 2024, Anthropic published the Model Context Protocol. [1] It is a specification document, a written agreement that says: here is how AI models and external tools should talk to each other, including the format for announcing what equipment is available, how a tool request should be structured, and how results get sent back.

In December 2025, Anthropic donated the protocol to the Agentic AI Foundation, a directed fund under the Linux Foundation, co-founded by Anthropic, Block, and OpenAI with support from Google, Microsoft, AWS, Cloudflare, and Bloomberg. This move formally made the protocol a community standard rather than a vendor-controlled one. [1]

The MCP server is the actual localized training documentation and physical kitchen setup that adheres to the global standard. It sits between the AI model and a vendor’s underlying product, translating core API capabilities into a set of labeled stations and specific tools the model can recognize and use at runtime.

A new employee — the AI — walks in. Because the store follows the standardized system, they know how to onboard efficiently. They see the labels, understand the specific tools, and can immediately start building the burger without needing a decade of experience in that specific building.

What Actually Happens When You Ask Claude a Question

When you first connect an AI client to an MCP server, there is a handshake. The client sends an initialize request. The server responds with a manifest: a structured list of every tool in the kitchen, with each one’s name, description, and the inputs it expects. It all happens over a protocol called JSON-RPC 2.0. [2]

That entire manifest gets loaded into the AI model’s context on every request. Not just once at setup. Every single message. Every tool, every label, every schema — handed to the model before it reads a single word of what you asked.

Tool definitions typically run anywhere from a few hundred to roughly a thousand tokens each, depending on schema complexity. [3] The model reads what you asked, looks at the available tools, decides which ones it needs. It then produces a structured request. That request travels to the MCP server. The server runs the underlying code, gets the result from the vendor’s API, and sends it back. The result re-enters the model’s context, and the model writes a response in plain English.

Not All Kitchens Are Equal

When a vendor announces “we have an MCP,” they mean they have an MCP server. That statement, on its own, tells you almost nothing about whether their AI integration is actually any good. Your corner bodega may tell you they have an operating manual... and then show you two grease-stained pages crumpled in a corner. Your bodega is not McDonald’s.

Three things determine whether an MCP server actually performs at the level the marketing implies: whether the kitchen is fully stocked, whether the labels make sense, and whether the stations are designed for service. Each fails independently, and each costs you in a different way.

The next piece in this series walks through those three failure modes, why tool count is a dance and not a number to maximize, and why a single well-designed server-side tool can collapse what would otherwise be fifty round trips into one.

→ Continue to Piece 2: The Three Ways an MCP Server Fails You (And Why Tool Design Is Everything)

References

1. Wikipedia — “Model Context Protocol” — https://en.wikipedia.org/wiki/Model_Context_Protocol — 2025
2. Anthropic / modelcontextprotocol.io — “Specification” — https://modelcontextprotocol.io/specification/2025-11-25 — 2025
3. MCP Playground — “MCP Token Counter: Why Your Tools Are Silently Eating Your Context Window” — https://mcpplaygroundonline.com/blog/mcp-token-counter-optimize-context-window — 2026
4. Attention Docs — “Attention MCP Server” — https://docs.attention.com/mcp/overview — 2026