mcp-server-design（MCP server 设计）

MCP Server Design

Overview

This skill provides best practices for designing MCP (Model Context Protocol) servers that work effectively with LLM agents. The key insight: design for agents, not automation. LLMs are human-like thinkers, not API consumers.

Core Philosophy

Design for Agents, Not Automation

Traditional API design optimizes for programmatic access with granular endpoints. MCP tool design should optimize for how LLMs think and reason:

• LLMs are human-like thinkers: They understand intent, context, and purpose
• Tools should be tasks, not endpoints: Shape tools around what users want to accomplish
• Reduce cognitive load: Fewer, more purposeful tools beat many granular ones

The Three Pillars

1. Give everything ready: Provide complete, actionable information
2. Reduce effort: Minimize steps needed to accomplish tasks
3. Reduce paths: Limit decision branches the LLM must navigate

Tool Design Principles

1. Purpose-Built Tools Over Generic Wrappers

Anti-pattern: Wrapping every API endpoint as a tool

// Bad: Generic database tools
// src/tools/run-sql.ts
// src/tools/list-tables.ts
// src/tools/describe-table.ts

Best practice: Design tools around user tasks

// Good: Task-oriented tools
// src/tools/prepare-database-migration.ts
import { z } from "zod";
import type { ToolMetadata } from "xmcp";

export const schema = {
  description: z.string().describe("What database changes are needed"),
};

export const metadata: ToolMetadata = {
  name: "prepare-database-migration",
  description: "Design a database change with safety checks and reviewed migration plan",
};

// src/tools/analyze-slow-queries.ts
// src/tools/create-backup.ts

2. Minimize Tool Count

LLMs struggle with long tool lists. Each additional tool:

• Increases selection confusion
• Adds tokens to every request
• Dilutes the purpose of each tool

Guidelines:

• Start with 5-10 core tools
• Add tools only when evals show they're needed
• Combine related operations when sensible

3. Name Tools for Purpose, Not Implementation

Tool names guide LLM behavior. The name should describe the task, not the mechanism.

4. Shape Tools Like Tasks

For complex flows, one tool per task beats one tool per step.

Anti-pattern: Breaking deployment into atomic operations

// Bad: Too many granular tools
// src/tools/create-branch.ts
// src/tools/commit-changes.ts
// src/tools/push-branch.ts
// src/tools/create-pr.ts
// src/tools/run-tests.ts
// src/tools/merge-pr.ts

Best practice: Single tool for the complete task

// src/tools/deploy-changes.ts
import { z } from "zod";
import type { ToolMetadata } from "xmcp";

export const schema = {
  description: z.string().describe("What changes are being deployed"),
  runTests: z.boolean().default(true).describe("Run test suite first"),
  autoMerge: z.boolean().default(false).describe("Auto-merge after tests pass"),
};

export const metadata: ToolMetadata = {
  name: "deploy-changes",
  description: "Deploy current changes through the complete workflow",
};

export default function deployChanges({ description, runTests, autoMerge }) {
  // Handles: branch creation, commit, push, PR, tests, merge
}

Decision Framework

When to Create a New Tool

Create a new tool when:

• Users frequently ask for this specific capability
• The task has clear boundaries and purpose
• Existing tools can't accomplish it cleanly
• Evals show the LLM selects incorrect tools for this task

Don't create a new tool when:

• An existing tool can handle it with minor parameter changes
• It duplicates functionality (combine instead)
• It's a rare edge case (handle with existing tools + guidance)

When to Combine Operations

Combine multiple operations into one tool when:

• They're almost always used together
• The intermediate results aren't useful alone
• Splitting them creates unnecessary decision points

Keep operations separate when:

• Users need granular control
• Intermediate results have standalone value
• Combining would create an overly complex tool

Balancing Granularity

Quick Reference

Tool Design Checklist

• Named for purpose, not implementation
• Description explains when to use it
• Parameters have clear descriptions
• Response includes next steps or context
• Tested with actual LLM requests

Red Flags

• More than 15 tools in a single server
• Tools named after HTTP methods or endpoints
• Generic tools that "can do anything"
• Tools that require multiple calls for a single task

Resources

references/

• design-principles.md - Extended examples, anti-patterns, and real-world case studies

For deeper exploration of these concepts, read the design principles reference document.