An MCP builder skill is a set of instructions that guides a coding agent through building a Model Context Protocol server, and the best-known one is the official mcp-builder skill from Anthropic. This guide explains how such a skill differs from an MCP server, walks through what the official skill tells the agent to do, compares the other MCP skills in the directory, and gives a plain rule for choosing between a skill and a server.
Descriptions come from the skills' own SKILL.md files, the Model Context Protocol documentation and the directory's records. The editorial team did not build a server with each skill, so nothing here ranks the output.
Skill versus server: what MCP skills are for
The Model Context Protocol documentation describes MCP as an open-source standard for connecting AI applications to external systems, including data sources, tools and workflows. It compares the protocol to a USB-C port for AI applications. A server is the thing that sits on the other end of that port and exposes capabilities.
A skill is different in kind. It is a folder with a SKILL.md that an agent reads when the task matches its description. Anthropic's plugin documentation puts it plainly: skills, commands and agents enter context as instructions that shape what Claude does with tools it already has. They do not add a connection by themselves.
That gives MCP skills three jobs:
- Build a server. The agent follows a design and testing workflow while it writes the server.
- Operate or configure one. The skill explains how to register, wire up or call an existing server.
- Wrap an API. The skill converts an existing interface, such as an OpenAPI spec, into a server.
The MCP topic lists all of these. The post on how to create a Claude skill covers writing your own.
The official MCP builder skill
anthropics/mcp-builder lives in the anthropics/skills repository. Its frontmatter names the skill and describes it as a guide for creating high-quality MCP servers, for use when integrating external APIs or services in Python with FastMCP or in Node and TypeScript with the MCP SDK. The directory records an Apache-2.0 licence and a pass result from its static check. The frontmatter itself says the complete terms are in a LICENSE.txt in the skill folder. The repository README separately states that most skills are Apache-2.0 while the document skills are source-available.
The skill folder holds SKILL.md, a LICENSE.txt, a reference/ folder and a scripts/ folder. The reference folder includes MCP best practices, a Python guide, a TypeScript guide and an evaluation guide. The SKILL.md describes four phases:
- Research and planning. Decide between broad API coverage and specialized workflow tools, with a preference for broad coverage when unsure. Name tools consistently with action-oriented prefixes, keep descriptions concise, support filtering and pagination, and study the protocol documentation. TypeScript is recommended, with streamable HTTP for remote servers and stdio for local ones.
- Implementation. Set up the project, write shared pieces such as API clients, error handling and formatters, then add each tool with input and output schemas and tool annotations like read-only and destructive hints.
- Review and test. Check code quality, types and consistency, build the project, and try it with the MCP Inspector, which the skill launches through
npx @modelcontextprotocol/inspector. - Evaluations. Write ten realistic, read-only questions in a specified XML format that test whether a model can actually use the server to answer real tasks.
The evaluation phase is the part most easily skipped when you build by hand, and it is the reason to prefer a skill for this job. The skill treats the quality of a server as how well it lets a model finish real tasks, not how many endpoints it covers.
The skill ships in the example-skills plugin of the anthropics/skills marketplace, so the Claude Code route is /plugin marketplace add anthropics/skills, followed by installing example-skills@anthropic-agent-skills. That installs every skill in the bundle. The Claude Code page lists the single-skill routes, and the guide to plugins and marketplaces covers scopes and review.
Other MCP skills in the directory
The directory's MCP topic has hundreds of skills, from builders to skills that merely mention MCP. These are the ones that matter for building, wrapping or driving servers:
| Skill | What it does | Licence |
|---|---|---|
| anthropics/mcp-integration | Explains bundling MCP servers in a Claude Code plugin: config files, stdio, SSE, HTTP and WebSocket types, authentication | Apache-2.0 |
| mcp-use/mcp-builder | Builds, debugs and migrates TypeScript MCP servers and MCP Apps with the mcp-use framework, trusting the installed package's types | Apache-2.0 |
| mcp-use/openapi-to-mcp | Turns an OpenAPI or Swagger spec into a server, mapping each operation to a tool, then wiring auth, testing and deploying | Apache-2.0 |
| shareai-lab/mcp-builder | Guide to tool, resource and prompt design with TypeScript or Python workflows | Apache-2.0 |
| jnmetacode/mcp-builder | Methodology for designing, building and testing servers, written in Chinese | MIT |
| codewhale-hq/mcp-builder | Builds and debugs servers for the Codewhale agent, including stdio and HTTP or SSE transports | MIT |
| sickn33/mcp-builder | Description nearly identical to the official skill's | Apache-2.0 |
| yusufkaraaslan/skill-builder | Uses the Skill Seekers MCP tools to turn docs, repos, PDFs or videos into packaged skills | MIT |
All of these show a pass result from the directory's static check at the time of writing. Choosing among them comes down to the framework. The official skill is framework-neutral and works with the SDKs directly. The mcp-use skills assume that library. The Codewhale skill is tied to one agent. The sickn33 listing sits in a larger collection and describes the same job as the official skill in nearly the same words, so install one or the other, not both.
Two further listings show a different pattern, where a skill drives an existing server rather than building one. htmlstreamofficial/preline-mcp adds Preline UI components by querying the Preline MCP server, and mksglu/context-mode-3 routes large command output through context-mode tools. The directory records no recognized licence for either (NOASSERTION), so read the repository's licence file before reuse.
Should you use a skill or a server?
Use this rule of thumb, drawn from what each one is.
| You need | Reach for | Why |
|---|---|---|
| The agent to call a live API, database or app | An MCP server | Only a server provides the tool connection |
| The agent to follow a procedure or house style | A skill | Instructions load on demand and need no process |
| To build a server with consistent design | An MCP builder skill | It carries naming, error and evaluation conventions |
| To use a server well | A skill that documents it | It teaches when and how to call the tools |
| Both | A plugin | One install can bundle skills and servers |
Cost differs too. Anthropic's documentation says the name and description of each skill stay in context on every turn while its full text loads only on use, and that MCP servers run as processes alongside the session. Other projects warn about the token cost of enabling many MCP tool descriptions at once. The ECC README, for example, cautions against turning on every server together. A skill that only documents a tool you rarely use is cheap, and a server you rarely use can still add turns of overhead.
What to check before you rely on one
- The generated server is your code. A builder skill writes a program that will hold API keys and call external services. Review what it requests and where credentials are stored.
- Read the skill's scripts. The official skill's
scripts/folder runs on your machine. The MCP Inspector command runs throughnpx, which downloads a package. - Match the SDK. Skills can lag SDK releases. Check the current protocol documentation for details such as transports and authentication.
- Mind the licence on copies. Several listings are independent works and some are close copies. State the licence you actually use, not the one on the first listing you found.
- Prefer pass results, then read anyway. The static check is a first filter and does not review the generated code.
For a wider view of how skills, plugins and subagents fit together, the subagents topic and the guide to Claude Code plugins are the next reads.
Frequently asked questions
What is the mcp-builder skill?
It is an official skill in the anthropics/skills repository that guides an agent through designing, building, testing and evaluating a Model Context Protocol server. It covers TypeScript with the MCP SDK and Python with FastMCP, and it ends with a set of evaluation questions that check whether a model can use the finished server.
What is the difference between an MCP skill and an MCP server?
An MCP server is a running program that exposes tools, data or prompts to an AI application over the Model Context Protocol. A skill is a folder of instructions the agent loads on demand. A skill can teach the agent to build or operate servers, but it does not itself provide a new tool connection.
Do I need a skill to build an MCP server?
No. You can build one from the protocol documentation and an official SDK. A skill packs design conventions, a workflow and checks into instructions the agent follows, which saves you repeating them in every prompt. It helps most when the agent writes much of the server for you.
Which licence does the official mcp-builder skill use?
The directory records Apache-2.0 for anthropics/mcp-builder. The skill's own frontmatter points to a LICENSE.txt in its folder for the complete terms, and the repository README says most skills are Apache-2.0 while the document skills are source-available.
Are the other mcp-builder skills copies of the official one?
Some are independent, such as the mcp-use skill for that TypeScript framework and the Chinese-language methodology skill. At least one listing has a description almost identical to Anthropic's. Compare the SKILL.md files and licences before installing more than one.