Chapter 11: Action Plan and Next Steps

You now have the knowledge to build, test, and deploy MCP servers. This chapter gives you a concrete checklist, project ideas to practice with, and pointers to community resources.

Your MCP Server Checklist

Work through this list to go from idea to production:

Phase 1: Build

Install the MCP Python SDK: pip install "mcp[cli]"
Define your server with FastMCP("your-server-name")
Implement at least one tool with @mcp.tool()
Write clear docstrings for every tool (the AI reads them)
Test locally with mcp dev server.py
Verify tools appear correctly in the MCP inspector
Call each tool from the inspector and check the output

Phase 2: Connect

Connect to Claude Desktop (add entry to claude_desktop_config.json)
Restart Claude Desktop and verify tools appear
Have a conversation that triggers tool use — confirm it works end to end
Connect to Claude Code (create .mcp.json in your project)
Run /mcp in Claude Code to verify the connection

Phase 3: Harden

Add error handling to all tools (timeouts, validation, graceful failures)
Log to stderr only — never to stdout with stdio transport
Add input validation for all tool parameters
If using resources: validate URI paths against a safe root
Test edge cases: empty inputs, very large inputs, missing data

Phase 4: Deploy (for HTTP servers)

Choose transport: HTTP/SSE or Streamable HTTP
Add authentication (OAuth or API key)
Add rate limiting
Set up a reverse proxy (nginx/Caddy) with TLS
Add a health endpoint
Containerize with Docker
Add structured logging

Project Ideas

The best way to learn MCP is to build something you will actually use.

Beginner Projects

Internal knowledge base search Connect your team’s documentation or wiki to Claude. Implement a search tool that queries your docs and returns relevant passages.

Calendar and task tools Expose your calendar and task list to Claude. Tools: list_today_events, create_task, mark_complete. Claude becomes your personal assistant.

File system browser A server that exposes a safe subset of your file system as resources. Claude can read files, list directories, and search for content.

Intermediate Projects

Database assistant Connect Claude to a PostgreSQL or SQLite database. Implement tools for querying data, summarizing tables, and generating reports. Add safety constraints (read-only, no DROP statements).

Git tools Expose git operations: git_log, git_diff, git_blame, git_status. Claude can help you understand what changed and why.

API integration hub Build a single MCP server that wraps multiple internal APIs (Jira, Confluence, PagerDuty). Claude can query tickets, read documentation, and check incident status in one conversation.

Advanced Projects

Agentic code reviewer A server that uses sampling to have Claude review its own analysis — a meta-reasoning loop for automated code quality checks.

Multi-step data pipeline A server that orchestrates a multi-step ETL pipeline, reporting progress with ctx.report_progress() and allowing Claude to monitor and intervene.

MCP proxy A server that connects to other MCP servers and exposes a unified tool namespace — useful for composing multiple specialized servers.

Community Resources

Official Resources

MCP Specification: modelcontextprotocol.io — the authoritative protocol documentation
Python SDK: github.com/modelcontextprotocol/python-sdk — source, examples, and changelog
Server Registry: modelcontextprotocol.io/servers — community-built servers for databases, APIs, and tools

SDKs in Other Languages

The MCP ecosystem has official SDKs for:

TypeScript/JavaScript — the most widely used SDK
Python — used throughout this book
Kotlin — for JVM environments
Swift — for Apple platforms
Go, Rust, C# — community SDKs

Learning by Example

The best way to understand advanced patterns is to read existing MCP servers. The official server registry links to dozens of open-source implementations. Study how they structure their tools, handle errors, and manage state.

What to Build Next

Now that you have the fundamentals, here are three directions to explore:

1. Go deeper on agents Combine your MCP server with the Claude API’s tool use to build fully autonomous agents. The agent loop from Chapter 9 is your starting point. Add memory, planning, and multi-step reasoning.

2. Build for your team The highest-value MCP servers are the ones that give your team superpowers. What does your team search for repeatedly? What manual processes could be automated? Build the server that changes how your team works.

3. Contribute to the ecosystem If you build something generally useful, consider publishing it. The MCP server registry welcomes community contributions, and well-built servers get adopted widely.

Final Thoughts

MCP is a young protocol, but it is already reshaping how developers integrate AI into their workflows. The tooling is maturing quickly, the ecosystem is growing, and the pattern of building reusable, portable AI integrations is clearly the direction the industry is moving.

You now have everything you need to be part of that ecosystem. Start small, test thoroughly, and build something that makes your own work meaningfully better. That is the best foundation for anything more ambitious.

← Chapter 10: Advanced Topics

Table of Contents

Gaëlle Candel