Clicked Gallery

What is MCP (Model Context Protocol)?

Highlighted from a real engineering doc. Explained by Clicked.

Used in a sentence

Engineering Notes · AI Systems

The accounting suite now ships an MCP server, so any AI assistant that speaks the protocol can pull a client's ledger without a custom integration.

The reader highlighted one word in the docs. Clicked explained the technical term β€œMCP” in simple terms:

Explained in three depths

Same facts, different vibe β€” Slang mode 😎

The Clicked way

●○○

Overview

MCP, the Model Context Protocol, is an open rulebook that sets one shared way for an AI app to connect to an outside tool. The AI app is the program with a model inside, such as a chat assistant. A tool is any outside service the app might use, such as Google Drive or a company database. Each tool gets a server: one small program that speaks the rulebook on one side and the tool's own language on the other. Each AI app gets a client: one piece that speaks the rulebook on the app's side. Because both follow the same rules, any client can use any server. Without a shared rulebook, 10 AI apps and 20 tools would need 200 separate connections, 10 times 20. With MCP they need 10 clients and 20 servers, 30 pieces of software, and every app can reach every tool.
●○○

Overview

MCP (Model Context Protocol) is the shared rulebook that lets any AI app hook up to any outside tool without a custom job for each pair. The AI app is the one with the model in it, like a chat assistant. The tool is Drive, Slack, your company's database, whatever the app needs to reach. Say 5 AI apps want to use 8 tools. Build a bespoke link for every pair and that's 40 jobs, 5 times 8. With MCP, each tool gets one server that speaks the rulebook, each app gets one client that speaks it too, and that's 13 jobs, 8 plus 5, for the same 40 links. "Now supports MCP" just means someone built one of the 13. 😎

A quick take β€” often all you need.

●●○

Detail

MCP, the Model Context Protocol, is an open rulebook for how an AI app connects to an outside tool, so that a tool connected once can be used by any AI app that follows the rules. The AI app is the program with a model inside, and the software around the model that carries out its requests is the harness. A tool is any outside service: a file store, a chat service, a database. MCP itself is only the rulebook, and it fixes the shape of three exchanges. A tool must answer the question "what do you offer?" with a list, each entry giving a name, a description and the details needed. It must accept a request in a set shape, and it must reply in a set shape. Two pieces of software follow those rules. A server is written for one tool and sits in front of it, speaking MCP outward and the tool's own language inward. The Google Drive server, for instance, offers "search files" and "read file", and does the real work inside Drive. A client sits inside each AI app, asks every server what it offers, hands the list to the model, and passes the model's requests through. The model never talks to a server. The model asks for a tool by name with the details filled in, which is tool calling, and the harness sends that request to the right server and pastes the reply into the conversation. A server can offer three kinds of thing: tools to run, data to read, and ready-made prompts. Anthropic published the rulebook on 25 November 2024 with six ready-made servers, and because the rules are open, anyone can write a server for any tool. MCP is not an API. An API is one service's own rules for being talked to, and a server usually sits in front of one and translates. The risk comes with the convenience: a server is someone else's program running with whatever access you gave it, and what it sends back is text the model will read and may act on.
●●○

Detail

MCP (Model Context Protocol) is a rulebook that settles how an AI app and an outside tool talk, so the tool gets wired up once and every app can use it. The AI app is the one with the model inside, and the software round the model that runs its errands is the harness. The tool is anything living outside: Drive, Slack, a database. The rulebook on its own does nothing. It only says what the chat between app and tool has to look like. Opening line, every time: what have you got? Back comes a list, each item with a name, a line on what it does and the blanks to fill. From then on, requests and replies in one fixed shape. Two bits of software do the actual talking. The server stands guard over one tool and speaks both languages, MCP on the outside and Drive-speak on the inside, and it's the server that really opens the file. The client lives in the AI app, collects the list from every server and gives the list to the model. The model and the server are never in the same room. The model fills in a request for a tool by name, and that's tool calling. The harness ferries the request out and the reply home. A server can offer 3 things: stuff to do, stuff to read, and ready-written prompts. Not an API, either. An API is one service's house rules, and a server mostly stands in front of one as the go-between. The catch: a server is a stranger's program holding the keys you handed over, and its replies are text the model reads and might act on. 😎

Want more? One click digs deeper.

●●●

Analogy

MCP, the Model Context Protocol, is one shared rulebook for how an AI app connects to any outside tool, so the tool is connected once and every app can use it. The Common App for US colleges is the same idea. A student used to fill in a separate application for every college, each with its own questions and its own layout. Now the student fills in one standard application, and more than 1,200 colleges accept it. The standard form is the rulebook. The student's one application, written to that form, is the server: the tool describing itself once in the agreed shape. Each college that reads the form is an AI app with a client. The form does not decide whether the student gets in; each college still reads and judges, just as each AI app still decides what to do with a tool. Where it breaks: the Common App is run by one organisation, while MCP is an open rulebook anyone can build to without asking. And colleges add their own extra questions, while a server offers only what it lists.
Deepest level reached.
●●●

Analogy

MCP (Model Context Protocol) is the shared rulebook that lets any AI app connect to any outside tool, because both sides build to the same rules. The shipping container is the same deal. Before it, every cargo got loaded by hand, differently for every ship, truck and train. Then everyone agreed on one steel box, 20 or 40 feet long, and now a crane in any port lifts any shipper's goods without knowing or caring what's inside. The box is the rulebook. A shipper packing goods into a box is a tool wrapped in a server. A port with a standard crane is one of the AI apps, client and all. What's inside the box is the tool's own language, which the port never has to learn. Where it comes apart: a box only carries things, while a server also answers questions. And a box is just steel, while a server runs code with whatever access you gave it, so it matters whose box you let through the gate. 😎
Deepest level reached.

Unfamiliar concept? A real-world example makes it click β€” fresh analogies on tap.

AI explanations may contain errors · Not professional advice

Formal definition β€” The same term, explained the usual way

The Model Context Protocol (MCP) is an open specification, released by Anthropic on 25 November 2024, that standardises how an AI application (the host) connects to external data sources and tools. A host runs one client per connection; each client maintains a session with an MCP server, a program that exposes three primitives: tools (callable functions), resources (readable data) and prompts (reusable templates). Messages use JSON-RPC 2.0 over two transports, stdio for local processes and streamable HTTP for remote servers. The specification, SDKs in several languages and a repository of reference servers (initially Google Drive, Slack, GitHub, Git, Postgres and Puppeteer) are open source. MCP governs discovery and invocation between host and server; it does not define how the model selects tools, which remains the host's function-calling layer.

Want Clicked to explain terms like “MCP” directly in your browser β€” including on PDFs?

Add to Chrome β€” Free

50 free Explanations · No credit card required