Question 1

What is an MCP server?

Accepted Answer

A process that gives an AI host — Claude Desktop, Claude Code, Cursor, an agent you wrote — tools it can call, resources it can read and prompts it can offer, over the Model Context Protocol. The host connects, asks what the server has, and the model calls into it. Every agent generator on this site consumes MCP servers; this page makes one.

Question 2

What goes in a server?

Accepted Answer

Three registries. Tools are functions the model calls, each with a name, a description it reads, a Zod input schema and optionally an output schema so results come back as structured data. Resources are things a client reads by URI — a static one like docs://readme, or a template like users://{id}/profile whose reader receives the variables. Prompts are reusable message templates the host exposes, usually as /name, whose arguments are always strings.

Question 3

stdio or Streamable HTTP?

Accepted Answer

stdio when a host spawns the server as a child process on the same machine — the common case for local tools, and what claude mcp add and a Claude Desktop config entry do. Streamable HTTP when the server runs somewhere else and hosts connect to a URL. The generated HTTP server is stateless: a fresh server and transport per request and no session id, which is the simplest shape to scale and to put behind a load balancer. Either way the server only ever logs to stderr, because on stdio stdout is the protocol.

Question 4

What do the tool annotations do?

Accepted Answer

They are hints to the client, not enforcement. readOnlyHint says the tool changes nothing, destructiveHint that it may destroy data, idempotentHint that repeating a call is harmless, openWorldHint that it reaches systems outside the server. A host may use them to decide what to auto-approve or what to run in parallel, so they should describe the handler honestly — the page flags a tool marked both read-only and destructive.

Question 5

Why does the template resource say list: undefined?

Accepted Answer

Because the SDK's ResourceTemplate constructor requires the list callback to be stated, even as undefined, so that forgetting resource listing cannot happen by accident. Return matching URIs from it when a client should be able to browse the template's instances; leave it undefined when the variables only make sense if the client already knows them.

Question 6

Why are names checked instead of fixed?

Accepted Answer

Because every name reaches the client verbatim. A tool name is what the model calls, a resource URI is what the client reads, a prompt name is the slash command the host shows, and the SDK throws on a duplicate registration. A silent rename would hand you a server whose tool answers to a name you never wrote down, so a bad one is reported and left out — and each registration's file is named after it, so the tree reads as the server's API.

Question 7

Does this generator send my definition anywhere?

Accepted Answer

No. The page runs entirely in your browser, needs no account, and never asks for a credential — the env var the server requires is emitted as a name in .env.example. The download is a .zip built client-side.

Question 8

Is this an official Model Context Protocol tool?

Accepted Answer

No. It is a free tool built by Nodlume against the public TypeScript SDK, pinned to the 1.x API its spellings were checked against. The SDK itself, and the last word on anything here, lives at modelcontextprotocol.io and the @modelcontextprotocol/sdk package.

MCP server generator

One file per registration

Names are never auto-fixed

Runs in the Inspector

How to build a custom MCP server

Name the server and pick a transport

Define tools

Add resources and prompts

Build, inspect and install

What the download contains

MCP servers, answered

Build the agent that calls it