mcp-builder

スター11

フォーク3

更新日2026年3月18日 12:40

Guide for building high-quality MCP (Model Context Protocol) servers and tool suites. Use this whenever the user asks to build, design, refactor, review, or evaluate an MCP server, expose an API through MCP, or improve tool naming, schemas, pagination, auth, or error handling.

インストール

Codex または Claude でインストールこの Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。

ソース

Threat-Vector-Security

Threat-Vector-Security/guardian-agent

GitHub リポジトリを開く Creator のリポジトリを見る

ダウンロード

MCP Builder

Build MCP servers so they are easy for agents to discover, reason about, and use safely.

Workflow

Understand the target API or system first.
Choose the transport intentionally.
Design the tool surface for agent discoverability, not just endpoint parity.
Implement shared auth, error handling, formatting, and pagination helpers before adding many tools.
Test the server with realistic multi-step prompts, not only unit tests.

Design Principles

Prefer clear verb-first or provider-prefixed tool names.
Make tool descriptions concrete and concise.
Use structured schemas with field constraints and examples.
Return structured data whenever the SDK supports it.
Keep read-only and mutating tools clearly separated.
Write actionable errors that tell the agent what to do next.

Implementation Guidance

Prefer TypeScript unless the repo already has a strong Python MCP pattern.
Start with the most common read-only tools, then add mutating operations.
Add pagination, filtering, and scoping so tools do not dump oversized payloads.
Model auth once in a shared client layer.
Use annotations and metadata that help approval and policy systems classify risk.

Review Checklist

Can an agent find the right tool from the name alone?
Are input and output schemas specific enough to prevent ambiguous calls?
Do error messages explain the likely fix?
Are mutating tools obviously distinct from inspection tools?
Can common user tasks be solved without custom scripting around the server?

Read references/mcp-server-checklist.md when you need a compact implementation checklist.

Gotchas

Do not mirror raw endpoint naming when it makes tool discovery worse for agents.
Do not expose large unfiltered list endpoints without pagination, filters, or scoping.
Do not mix read-only and mutating behavior behind one ambiguous tool.

このリポジトリの他の Skills

同じリポジトリ

skill-creator

Threat-Vector-Security/guardian-agent

Guide for creating new skills, improving existing skills, and evaluating skill quality. Use this whenever the user wants to turn a workflow into a skill, revise a skill, improve triggering behavior, or add evals and benchmarks for a skill.

2026-04-1911

code-review

Threat-Vector-Security/guardian-agent

Use when reviewing a diff, PR, or patch for bugs, regressions, missing tests, architecture risk, security risk, or reviewability concerns before merge.

2026-04-1311

context-engineering

Threat-Vector-Security/guardian-agent

Use when starting a repo-grounded task, switching subsystems, or when output quality drops and the agent needs tighter project context from rules, specs, relevant files, examples, and current errors.

2026-04-1311

source-driven-development

Threat-Vector-Security/guardian-agent

Use when implementing or reviewing framework-specific code that should be verified against current official docs, versioned guidance, or source-cited examples rather than memory.

2026-04-1311

coding-backend-orchestration

Threat-Vector-Security/guardian-agent

Use when the user explicitly asks Guardian to launch or check an external coding assistant such as Codex, Claude Code, Gemini CLI, or Aider.

2026-04-1111

coding-workspace

Threat-Vector-Security/guardian-agent

Use when the request is about a repo, codebase, implementation, bugfix, or backend-owned coding session that should stay anchored to the active workspace.

2026-04-1111

name	mcp-builder
description	Guide for building high-quality MCP (Model Context Protocol) servers and tool suites. Use this whenever the user asks to build, design, refactor, review, or evaluate an MCP server, expose an API through MCP, or improve tool naming, schemas, pagination, auth, or error handling.

MCP Builder

Build MCP servers so they are easy for agents to discover, reason about, and use safely.

Workflow

Understand the target API or system first.
Choose the transport intentionally.
Design the tool surface for agent discoverability, not just endpoint parity.
Implement shared auth, error handling, formatting, and pagination helpers before adding many tools.
Test the server with realistic multi-step prompts, not only unit tests.

Design Principles

Prefer clear verb-first or provider-prefixed tool names.
Make tool descriptions concrete and concise.
Use structured schemas with field constraints and examples.
Return structured data whenever the SDK supports it.
Keep read-only and mutating tools clearly separated.
Write actionable errors that tell the agent what to do next.

Implementation Guidance

Prefer TypeScript unless the repo already has a strong Python MCP pattern.
Start with the most common read-only tools, then add mutating operations.
Add pagination, filtering, and scoping so tools do not dump oversized payloads.
Model auth once in a shared client layer.
Use annotations and metadata that help approval and policy systems classify risk.

Review Checklist

Can an agent find the right tool from the name alone?
Are input and output schemas specific enough to prevent ambiguous calls?
Do error messages explain the likely fix?
Are mutating tools obviously distinct from inspection tools?
Can common user tasks be solved without custom scripting around the server?

Read references/mcp-server-checklist.md when you need a compact implementation checklist.

Gotchas

Do not mirror raw endpoint naming when it makes tool discovery worse for agents.
Do not expose large unfiltered list endpoints without pagination, filters, or scoping.
Do not mix read-only and mutating behavior behind one ambiguous tool.