Agent Skills: MCP Server Development Guide

MCP Server Development Guide

Process

High-Level Workflow

Build an MCP server in four phases:

Phase 1: Deep Research and Planning

1.1 Understand Agent-Centric Design Principles

Before implementation, design tools around agent workflows:

Build for Workflows, Not Just API Endpoints:

• Do not mirror every API endpoint by default. Group endpoints into workflow tools when one user task requires multiple API calls.
• Consolidate related operations (e.g., schedule_event that both checks availability and creates event)
• List the top user workflows and map each tool to one workflow.

Optimize for Limited Context:

• Return only fields needed for the task by default.
• Return high-signal information, not exhaustive data dumps
• Provide "concise" vs "detailed" response format options
• Default to human-readable identifiers over technical codes (names over IDs)
• Add pagination, field selection, or truncation for large responses.

Design Actionable Error Messages:

• Error messages include the failed field or operation, the cause, and the next valid action.
• Suggest specific next steps: "Try using filter='active_only' to reduce results"
• Never return raw provider errors without a tool-level explanation.

Follow Natural Task Subdivisions:

• Tool names should reflect how humans think about tasks
• Group related tools with consistent prefixes for discoverability
• Design tools around natural workflows, not just API structure

Use Evaluation-Driven Development:

• Create realistic evaluation scenarios early
• Let agent feedback drive tool improvements
• Prototype quickly and iterate based on actual agent performance

1.2 Study MCP Protocol Documentation

Fetch the MCP protocol documentation:

Fetch: https://modelcontextprotocol.io/llms-full.txt

This document contains the complete MCP specification. Note: the URL may change as the spec evolves — if the fetch fails, search for the current MCP docs URL.

1.3 Study Framework Documentation

Load and read the following reference files:

• MCP Best Practices: 📋 View Best Practices - Core guidelines for all MCP servers

For Python implementations, also load:

• Python SDK Documentation: Fetch https://py.sdk.modelcontextprotocol.io/
• 🐍 Python Implementation Guide - Python-specific best practices and examples

For Node/TypeScript implementations, also load:

• TypeScript SDK Documentation: Fetch https://ts.sdk.modelcontextprotocol.io/
• ⚡ TypeScript Implementation Guide - Node/TypeScript-specific best practices and examples

1.4 Study API Documentation

To integrate a service, read the API documentation for the selected workflows. Scope endpoint and parameter coverage to those workflows, but read the cross-cutting concerns (auth, rate limits, error handling, data models) in full — they apply across every tool you expose:

• Official API reference documentation
• Authentication and authorization requirements
• Rate limiting and pagination patterns
• Error responses and status codes
• Available endpoints and their parameters
• Data models and schemas

Search or fetch documentation only for gaps that block the selected workflow.

1.5 Create an Implementation Plan

Create a plan with these sections:

Tool Selection:

• List the most valuable endpoints/operations to implement
• Prioritize tools that enable the most common and important use cases
• Consider which tools work together to enable complex workflows

Shared Utilities and Helpers:

• Identify common API request patterns
• Plan pagination helpers
• Design filtering and formatting utilities
• Plan error handling strategies

Input/Output Design:

• Define input validation models (Pydantic for Python, Zod for TypeScript)
• Design consistent response formats (e.g., JSON or Markdown), and configurable levels of detail (e.g., Detailed or Concise)
• Plan for large-scale usage (thousands of users/resources)
• Implement character limits and truncation strategies (e.g., 25,000 tokens)

Error Handling Strategy:

• Plan graceful failure modes
• Define each error shape with: error code, human-readable cause, retryability, and next valid action.
• Consider rate limiting and timeout scenarios
• Handle authentication and authorization errors

Phase 2: Implementation

After the plan exists, implement the shared infrastructure before individual tools.

2.1 Set Up Project Structure

For Python:

• Create a single .py file or organize into modules if complex (see 🐍 Python Guide)
• Use the MCP Python SDK for tool registration
• Define Pydantic models for input validation

For Node/TypeScript:

• Create the project structure from ⚡ TypeScript Guide
• Set up package.json and tsconfig.json
• Use MCP TypeScript SDK
• Define Zod schemas for input validation

2.2 Implement Core Infrastructure First

To begin implementation, create shared utilities before implementing tools:

• API request helper functions
• Error handling utilities
• Response formatting functions (JSON and Markdown)
• Pagination helpers
• Authentication/token management

2.3 Implement Tools Systematically

For each tool in the plan:

Define Input Schema:

• Use Pydantic (Python) or Zod (TypeScript) for validation
• Include constraints (min/max length, regex patterns, min/max values, ranges)
• Describe field purpose, accepted values, and defaults
• Include diverse examples in field descriptions

Write Tool Docstrings/Descriptions:

• One-line summary of what the tool does
• Purpose and functionality
• Explicit parameter types with examples
• Return type schema
• Usage examples (when to use, when not to use)
• Error handling documentation, which outlines how to proceed given specific errors

Implement Tool Logic:

• Use shared utilities to avoid code duplication
• Follow async/await patterns for all I/O
• Implement the planned error shapes
• Support multiple response formats (JSON and Markdown)
• Respect pagination parameters
• Check character limits and truncate appropriately

Add Tool Annotations:

• readOnlyHint: true (for read-only operations)
• destructiveHint: false (for non-destructive operations)
• idempotentHint: true (if repeated calls have same effect)
• openWorldHint: true (if interacting with external systems)

2.4 Follow Language-Specific Best Practices

Load the matching guide before finishing implementation — 🐍 Python Implementation Guide or ⚡ TypeScript Implementation Guide — and follow its patterns: proper SDK tool registration (@mcp.tool / server.registerTool), strict input validation (Pydantic v2 model_config / Zod .strict()), full type coverage with no any, async/await for all I/O, module-level constants for shared values (e.g. CHARACTER_LIMIT, API_BASE_URL), and — for TypeScript — a working bun run build.

Phase 3: Review and Refine

After initial implementation:

3.1 Code Quality Review

Review the code for:

• DRY Principle: No duplicated code between tools
• Composability: Shared logic extracted into functions
• Consistency: Similar operations return similar formats
• Error Handling: All external calls have error handling
• Type Safety: Full type coverage (Python type hints, TypeScript types)
• Documentation: Every tool has summary, parameters, return shape, examples, and error behavior

3.2 Test and Build

Important: MCP servers are long-running processes that wait for requests over stdio/stdin or sse/http. Running them directly in the main process (e.g., python server.py or node dist/index.js) hangs it indefinitely.

Safe ways to test: run the server in tmux to keep it out of the main process, use a timeout (timeout 5s python server.py), or let the evaluation harness (Phase 4) manage the server subprocess directly for stdio transport.

• Python: verify syntax with python -m py_compile your_server.py and review imports.
• Node/TypeScript: run bun run build, confirm it completes without errors, and verify dist/index.js is created.

3.3 Use Quality Checklist

To verify implementation quality, load the appropriate checklist from the language-specific guide:

• Python: see "Quality Checklist" in 🐍 Python Guide
• Node/TypeScript: see "Quality Checklist" in ⚡ TypeScript Guide

Phase 4: Create Evaluations

After implementing the MCP server, create evaluations that test whether agents can answer realistic questions with the tools.

Load ✅ Evaluation Guide for complete evaluation guidelines.

4.1 Understand Evaluation Purpose

Evaluations test whether LLMs can effectively use your MCP server to answer realistic, complex questions.

4.2 Create 10 Evaluation Questions

Follow the process outlined in the evaluation guide:

1. Tool Inspection: List available tools and understand their capabilities
2. Content Exploration: Use READ-ONLY operations to explore available data
3. Question Generation: Create 10 complex, realistic questions
4. Answer Verification: Solve each question yourself to verify answers

4.3 Evaluation Requirements

Each question must be:

• Independent: Not dependent on other questions
• Read-only: Only non-destructive operations required
• Complex: Requiring multiple tool calls and deep exploration
• Realistic: Based on real use cases humans would care about
• Verifiable: Single, clear answer that can be verified by string comparison
• Stable: Answer won't change over time

4.4 Output Format

Create an XML file with this structure:

<evaluation>
  <qa_pair>
    <question>Find discussions about AI model launches with animal codenames. One model needed a specific safety designation that uses the format ASL-X. What number X was being determined for the model named after a spotted wild cat?</question>
    <answer>3</answer>
  </qa_pair>
<!-- More qa_pairs... -->
</evaluation>

Gotchas

• Running the server directly hangs the process. MCP servers block on stdio/stdin waiting for requests indefinitely. Never run python server.py or node dist/index.js in your main process. Use tmux to background the server, or use the evaluation harness which manages the server subprocess.
• Live URL references can go stale. The MCP protocol documentation URL (modelcontextprotocol.io/llms-full.txt) and the GitHub SDK README URLs may move between versions. If a fetch fails, search for the current URL rather than assuming the skill path is correct.
• TypeScript builds must be re-run after every source change. dist/index.js is the artifact the MCP runtime loads. Editing .ts files without running bun run build means the runtime is still running the old version.
• Tool annotations are hints, not enforced contracts. readOnlyHint: true does not prevent a tool from writing data — it only signals intent to the LLM. Enforce safety in the tool implementation itself.

Reference Files

Load only the resources required by the implementation language and phase.

| File | Load during | Covers | |---|---|---| | references/mcp_best_practices.md | Phase 1.3 (load first) | naming conventions, response formats (JSON vs Markdown), pagination, character limits/truncation, security and error-handling standards | | references/python_mcp_server.md | Phase 2 (Python) | server init, Pydantic models, @mcp.tool registration, working examples, quality checklist | | references/node_mcp_server.md | Phase 2 (TypeScript) | project structure, Zod schemas, server.registerTool, working examples, quality checklist | | references/evaluation.md | Phase 4 | question-creation and answer-verification guidelines, XML format, example Q&A, running evaluations |

Also fetch the live protocol/SDK docs as needed: MCP protocol spec (https://modelcontextprotocol.io/llms-full.txt), Python SDK (https://py.sdk.modelcontextprotocol.io/), TypeScript SDK (https://ts.sdk.modelcontextprotocol.io/).