skillmake
← marketplace
aitoolsha:101e939af858d34amanual

mcp-builder

Use when building a Model Context Protocol server that exposes a real external service to an LLM, so the agent designs the right tools and ships in Python (FastMCP) or Node/TypeScript with the official SDK.

Install confidence
curl --create-dirs -fsSL https://skillmake.xyz/i/mcp-builder -o ~/.claude/skills/mcp-builder/SKILL.md
Pinned content
sha:101e939af858d34a
Generated with
manual
Source
github.com

The file served at /api/marketplace/mcp-builder-101e939a/raw matches this hash. Inspect before install, then copy the command.

3,027 chars · ~757 tokens
---
name: mcp-builder
description: Use when building a Model Context Protocol server that exposes a real external service to an LLM, so the agent designs the right tools and ships in Python (FastMCP) or Node/TypeScript with the official SDK.
source: https://github.com/anthropics/skills/tree/main/skills/mcp-builder
generated: 2026-05-25T05:10:32.739Z
category: tool
audience: ai
---

## When to use

- Wrapping an existing API as an MCP server so a Claude or Cursor agent can call it natively
- Designing the tool surface for a new MCP server — picking which actions become tools, what the args look like, and how errors flow back
- Choosing between Python FastMCP and the Node/TypeScript MCP SDK for a new server
- Reviewing an existing MCP server for tool-design anti-patterns before broader adoption

## Key concepts

### Quality = task completion, not coverage

The skill measures an MCP server by how well an LLM can accomplish real tasks with it, not by how many endpoints it exposes. Five well-shaped tools beat 50 thin wrappers.

### Tool design as the primary surface

Each tool is a contract: name, args schema, return shape, error mode. The skill walks how to shape them so an LLM picks the right tool first try and recovers cleanly on failure.

### Python (FastMCP) vs TypeScript path

FastMCP is the fast track for Python services; the Node SDK is the path when the target codebase is JS/TS. The skill picks based on the host language, not personal preference.

### Local stdio vs remote transports

Stdio works for the developer's machine; HTTP/SSE transports unlock multi-user and remote deployment. Pick transport at design time, not retrofit later.

### Workflow-driven tool grouping

Tools that compose into a workflow (find → fetch → edit → confirm) reduce LLM round-trips. The skill steers tool boundaries to match how an agent actually completes work.

## API reference

```
npx skills add anthropics/skills --skill mcp-builder
```

Install the MCP server builder skill.

```
npx skills add anthropics/skills --skill mcp-builder
```

```
FastMCP / @modelcontextprotocol/sdk
```

Reference implementations the skill points at; pick based on the project language.

```
pip install fastmcp  # Python
npm install @modelcontextprotocol/sdk  # TypeScript
```

## Gotchas

- Don't expose every API endpoint; thin wrappers create discovery overhead for the LLM with no win for the user
- Tool args should describe intent, not URL parameters; 'find_issue_by_keyword' beats 'GET_issues_with_q_param'
- Errors must be parseable by an LLM; opaque stack traces or numeric codes leave the agent guessing at recovery
- Stdio transports don't survive multi-user setups; design for HTTP/SSE from the start if remote is on the roadmap
- MCP server names collide across registries; pick a unique, descriptive name early or you'll be renaming later

---
Generated by SkillMake from https://github.com/anthropics/skills/tree/main/skills/mcp-builder on 2026-05-25T05:10:32.739Z.
Verify against source before relying on details.

File: ~/.claude/skills/mcp-builder/SKILL.md