AI Artifacts & Generated Sites

MCP Server

Definition

An MCP server is the supply side of the protocol. It advertises capabilities in three kinds: tools, which are functions the model can call; resources, which are data it can read; and prompts, which are reusable templates. A server runs either as a local process the client starts and talks to over standard input and output, or as a remote service reached over HTTP much as a REST API is. Each tool declares a name, a description written for the model to read, and a JSON schema for its arguments, so the model knows what it may pass. Authentication belongs to the server: a remote one normally expects an API token, often issued through an OAuth flow. From the user's side almost none of this is visible — a new tool simply appears in the assistant and can be used.

Why It Matters

Without a server of some kind, an assistant can write you a file but can do nothing with it. In practice that is the difference between being handed some HTML and being handed a public URL. 99helpers runs an MCP server for this purpose, so an assistant connected to it can upload the page it has just written and return the live link in the same reply. That removes the download-then-upload step, which is where a fair number of decent prototypes quietly die. It also makes publishing repeatable rather than a manual drag of a file into a browser window.

How It Works

Client and server negotiate on connection, and the server sends its list of tools. When the model picks one, the client sends a call with arguments matching that tool's schema and waits for a result, which returns as text, structured data, or an error the model can read and respond to. A remote server checks credentials on every request, usually a bearer token in the Authorization header, and applies whatever limits belong to that account. Because the tool description is the only thing the model has to go on, a vague description is the usual reason a tool gets ignored or misused. Nothing happens on its own: a server acts only when it is called.

Real-World Example

Someone asks an assistant for a one-page event schedule. The assistant writes it, calls the publish tool on the connected server with the file and a chosen name, and replies with schedule-2026.99helpers.site already live. A follow-up asking to move the start times updates the same address instead of creating a second one, because the tool takes the existing name as an argument. Four manual steps become a sentence.

Common Mistakes

  • Confusing the server with the protocol — the protocol is the shared language, the server is one program that happens to speak it
  • Handing a server a long-lived credential with far more permission than the task needs, rather than scoping it to what the assistant is meant to do
  • Assuming a connected server lets the assistant act unprompted; a tool runs only when it is called inside a conversation

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 →