MCP Server Development

Build MCP servers that enable LLMs to accomplish real-world tasks through well-designed tools.

4-Phase Process

Research → Implement → Review → Evaluate

Load resources on-demand (progressive disclosure):

• Protocol: https://modelcontextprotocol.io/llms-full.txt
• Python SDK: https://raw.githubusercontent.com/modelcontextprotocol/python-sdk/main/README.md
• TypeScript SDK: https://raw.githubusercontent.com/modelcontextprotocol/typescript-sdk/main/README.md

Core Design Principles

Workflows over endpoints: Consolidate operations (e.g., schedule_event checks availability + creates) Optimize context: Return high-signal info, provide concise/detailed formats, use readable IDs Actionable errors: Guide to correct usage ("Try filter='active_only'"), suggest next steps Natural task flow: Name tools how humans think, group with consistent prefixes

Phase 1: Research & Planning

1.1 Core Docs

• Load protocol: https://modelcontextprotocol.io/llms-full.txt
• Read Best Practices

1.2 Framework (choose one)

• Python: Fetch SDK + read Python Guide
• TypeScript: Fetch SDK + read TypeScript Guide

1.3 API Research Study exhaustively: auth, rate limits, pagination, endpoints, errors, schemas

1.4 Plan

• Tool selection (most valuable endpoints)
• Shared utilities (requests, pagination, formatting, errors)
• I/O design (validation schemas, response formats, 25k char limits)
• Error strategy (graceful failures, actionable messages)

Phase 2: Implementation

2.1 Structure Python: single .py or modules | TypeScript: package.json + tsconfig.json

2.2 Infrastructure First Build shared utilities: API helpers, error handling, formatting, pagination, auth See: examples/basic-server.py or .ts

2.3 Tools (for each):

• Schema: Pydantic/Zod with constraints, clear descriptions
• Docs: Summary, purpose, params+examples, returns, usage, errors
• Code: Use utilities, async/await, multiple formats, respect limits
• Annotations: readOnlyHint, destructiveHint, idempotentHint, openWorldHint

See: examples/tool-with-context.py, error-handling.py

2.4 Best Practices Python: @mcp.tool(), Pydantic v2, type hints, async/await, constants TypeScript: registerTool(), Zod .strict(), strict mode, no any, Promise<T>

Phase 3: Review & Refine

3.1 Quality: Check DRY, composability, consistency, error handling, types, docs

3.2 Test: MCP servers are long-running. See testing.md Quick check: python -m py_compile server.py or npm run build

3.3 Checklist: Load from Python or TypeScript guide

Phase 4: Evaluations

Test LLM effectiveness with your server.

Process: Inspect tools → Explore data → Generate 10 questions → Verify answers

Requirements: Independent, read-only, complex, realistic, verifiable, stable

Format:

<evaluation>
  <qa_pair>
    <question>Complex question...</question>
    <answer>Verifiable answer</answer>
  </qa_pair>
</evaluation>

Complete guide: evaluation.md

Reference Library

Examples

Success Criteria

• Clear tool boundaries (single purpose)
• Efficient context (concise defaults)
• Educational errors (guide usage)
• Workflow optimization (consolidate steps)
• Production ready (errors, rate limits, logging)
• Agent effective (90%+ eval success)