Install
npx @modelcontextprotocol/inspectorWho it's for
Use mcp-builder when you want to create a high-quality MCP (Model Context Protocol) server so a model can call well-designed tools against an external API or service. The language can be Python (FastMCP) or Node/TypeScript (MCP SDK). It is a build guide, not a host for an unnamed server.
A better fit is when you are willing to do deep research and planning first, instead of writing interfaces immediately. The material writes TypeScript as the choice with better SDK support and better compatibility across multiple runtimes (for example MCPB). The material does not list agent products or a repository install path. The official guide page is
https://raw.githubusercontent.com/anthropics/skills/main/skills/mcp-builder/SKILL.md
First try in 3 minutes
This guide does not give a skill install command; do not invent a path on the first try. The recommended path starts at stage one: Deep Research and Planning. A high-quality MCP server has four main stages; the excerpt only fully writes the first-stage title, so the first try is planning only — do not write a server process.
Prerequisite: you already have an external API or service to connect. Low-risk action: list which queries on that service are read-only and which change data, and decide that remote uses Streamable HTTP. The summary also mentions stdio for local use, but the excerpt only writes Streamable HTTP for remote, so do not assume stdio arguments are already fully written.
Success signal: one page can state clearly "which tools the model gets, and what constraints the inputs use." Do not npm init an unplanned project first. After the plan is confirmed, move to implementation. On the TypeScript side you can use Zod for the input schema; fields need constraints and clear descriptions, and examples belong in the field descriptions. Confirm the directory before writing files.
Use cases
Case 1: Plan a remote server on the recommended TypeScript stack. Choose the MCP SDK, use Streamable HTTP for transport, Zod for inputs, and put constraints and examples in the descriptions. The summary also mentions outputSchema/structuredContent and annotations such as readOnlyHint and destructiveHint, but the excerpt does not write complete examples of those fields, so this case is done when the plan separates read-only tools from tools that change data, not when every annotation is already filled. Do not start a Python stack at the same time as the "main path"; Python FastMCP and Pydantic are parallel options, not a required double write.
Case 2: After implementation, compile and inspect. The TypeScript path is first npm run build to confirm it compiles, then MCP Inspector: npx @modelcontextprotocol/inspector. Inspector opens an inspection UI, so it has side effects; confirm you are testing the local or remote server you just built, then run. Done means build passes and Inspector can connect to the service you named. The excerpt does not write every Inspector flag, so do not invent connection parameters.
Boundaries
The four main stages start with deep research and planning; later the summary sketches implement then inspect/test, but the remaining stage titles are not expanded in the excerpt. Remote transport is written as Streamable HTTP. Input schema uses Zod or Pydantic. TypeScript check commands are npm run build and the official Inspector.
Do not invent an in-repo install path. The material does not give a client list, and does not say how to skip planning and publish directly. Beyond these written stages, language options, and check commands, this guide does not cover the rest.