AI Artifacts & Generated Sites

Model Context Protocol

Definition

The Model Context Protocol, usually shortened to MCP, was published by Anthropic in late 2024 and released openly for anyone to implement. It addresses a dull but expensive problem: before it, every assistant needed a bespoke integration for every tool. MCP defines that wiring once. A client, meaning the assistant application, connects to a server, asks what it offers, and receives a list of tools, resources and prompt templates described in JSON. Messages use JSON-RPC over either a local pipe or HTTP, and the protocol is versioned, so the two sides agree on a version they both understand. It is tied to no single model and no single vendor; implementations exist in several languages and a growing number of assistants support it.

Why It Matters

The value is plural rather than technical. One MCP server written once can be used by every client that speaks the protocol, which is how a small service becomes available inside several assistants without anyone writing a per-vendor adapter. For the person using it, this is the difference between an assistant that talks about your files and one that can fetch, change or publish them. It also keeps the boundary legible: the server decides what is exposed and enforces its own permissions, so connecting an assistant is not the same as handing over your account. Standards move, so check the current specification rather than a blog post from last year.

How It Works

A session opens with an initialisation exchange in which both sides state the protocol version and what they support. The client then requests the tool list, and the server replies with each tool's name, a description aimed at the model, and a JSON schema for its input. When the model chooses a tool, the client sends a call, the server runs it, and content comes back that the model can read — text, structured JSON, or an error. Local servers talk over standard input and output; remote ones use HTTP with an event stream for messages the server pushes, and authenticate each request with an API token. The protocol holds no opinion about what a tool does, so publishing a file and querying a database look identical on the wire.

Real-World Example

A freelancer connects their assistant to a hosting server that speaks MCP. They ask for a one-page invoice summary; the assistant writes it, calls the publish tool, and hands back invoices-q3.99helpers.site. Nothing in that flow was built for this particular assistant: the same server answers a different client in exactly the same way, and the freelancer's share of the work was ticking a box and pasting a token.

Common Mistakes

  • Treating MCP as a product rather than a specification — what you install is a client and a server, the protocol is just the agreement between them
  • Expecting every assistant to support it identically; the transports and the feature set differ, and both are still moving
  • Pasting a credential with full account access into a client you have not examined, on the assumption that the protocol limits what a server may expose

Related Terms

Put a file online in seconds

Drop in a document, an image, a page or a whole static website and share the link — free, with no build step and no server to set up.

Host a file free →