mcp-builder · MCP Server Builder

A build skill for high-quality MCP (Model Context Protocol) servers: wrapping external APIs and services as well-designed tools that LLMs can call.

A four-phase flow: research & planning (protocol spec, tool design principles) → implementation (TypeScript or Python tracks) → review & testing (MCP Inspector) → evaluation (10 real-world eval questions). Recommended stack: TypeScript, streamable HTTP for remote servers, stdio for local ones.

Example invocation: "Build an MCP server for our company's ticketing API."

Full brief

Positioning

mcp-builder is an end-to-end guide to building MCP servers from scratch: the quality bar is whether an LLM can use them to accomplish real tasks. It covers protocol research, tool design, implementation, testing, and evaluation — producing a shippable MCP server project.

Core capabilities

• Research & planning: studies the MCP protocol spec (sitemap → .md pages) and framework docs; weighs "full API coverage" against "workflow tools" as design approaches.
• Tool design standards: clear tool naming (consistent prefix + action verb), Zod/Pydantic input schemas, structured outputs via outputSchema, actionable error messages, annotations like readOnlyHint.
• Implementation: TypeScript (recommended) or Python (FastMCP) tracks; async I/O, pagination support, error-handling helpers.
• Review & testing: code quality checks (DRY, type coverage, consistent error handling); live testing with the MCP Inspector (npx @modelcontextprotocol/inspector).
• Evaluation: writes 10 independent, read-only, complex, verifiable evaluation questions per the eval guide — and solves each one to verify the answer.

Workflow

1. Research & planning: read the protocol spec and framework docs, list endpoints to implement
2. Implementation: project skeleton → shared infrastructure (auth, errors, pagination) → tool by tool
3. Review & testing: code review, Inspector sessions
4. Evaluation: 10 questions, output as XML

Inputs & outputs

Output: MCP server project (with tool implementations) and an evaluation XML file.

Boundaries with adjacent skills

• claude-api: the reference manual for building applications with the Claude API (consuming LLM capabilities).
• mcp-builder (this): the guide to building external tools (MCP servers) for LLMs (producing tools). An MCP server is one tool source a claude-api app can connect to.

Fit

• Wrapping internal company APIs as tools an LLM can call
• Building a standardized LLM access layer for a specific data source
• Quality evaluation and polish before shipping an MCP server

Before you start

• No API key needed; have the target service's API docs and auth method ready.
• Familiarity with TypeScript or Python helps; eval questions need a human check that answers are verifiable and stable over time.

mcp-builder is part of the Aiglade Skill library. Invoke it from the Aiglade chat box in plain language.

← All 155 skills