In partnership with
What's blocking you from building AI agents?
Every AI integration you've ever built follows the same pattern.
You want Claude to read your GitHub issues. You write custom code. It works for a week. GitHub changes their API. It breaks. You fix it. Repeat.
Now you want Claude to query your database. Different API, different authentication, different everything. You write more custom code.
This doesn't scale. Every tool needs custom integration. Every integration is fragile. Every update breaks something.
MCP fixes this.
What MCP does
MCP is a client-server protocol using JSON-RPC 2.0. Think of it as a standardized plugin system for AI.
The architecture is simple.
AI applications (clients) connect to external capabilities (servers) through a standard interface. Each server exposes specific tools. The AI can use any tool through the same protocol.
MCP servers expose three types of capabilities:
Resources are read-only data endpoints. A database server might expose
postgres://localhost/mydb/usersorpostgres://table/orders. The AI reads your database without you copying data manually.Tools are executable functions. A database server's
query_databasetool turns "show me last month's orders" into an actual SQL query that returns results.Prompts are reusable message templates. A data analysis prompt structures every analysis the same way, so Claude doesn't reinvent the format each time.
The protocol is stateless by design. Each request contains full context. Servers don't maintain sessions.
Build a working database server
Let's build a functional MCP server that connects AI to PostgreSQL. You'll be able to query your database and inspect schemas through natural language.
Step 1: Start with a new TypeScript project and install dependencies:
Step 2: Configure TypeScript for Node.js:
Step 3: Import the SDK and initialize the server:
Step 4: Register your tools. This tells AI clients what capabilities your server provides:
Step 5: Implement the tool execution logic:
Note: A critical rule that breaks most MCP servers: never write anything except JSON-RPC messages to stdout when using STDIO transport.
This is wrong and will break your server:
This is correct:
Any non-JSON output to stdout corrupts the message stream. The client can't parse it. Your server stops working. Always log to stderr.
Step 6: Connect the server to the STDIO transport:
Step 7: Test your server with the MCP Inspector before connecting to a real client:
The inspector provides a UI for testing tools manually and debugging issues. Essential during development.
Step 8: Configure Claude Desktop to use your server. Edit ~/Library/Application Support/Claude/claude_desktop_config.json:
Final step: Restart Claude Desktop. Your database server appears in the tools menu. Claude can now query your PostgreSQL database through natural language.
Making It production-ready
Your server works locally. Production systems need authentication, validation, and error handling.
Step 1: Add authentication checks before executing tools:
Step 2: Validate all inputs against JSON schemas before execution:
Step 3: Wrap external calls with proper error handling:
Step 4: For HTTP-based servers, add health checks:
Production checklist:
Validate all inputs with JSON Schema
Never log to stdout with STDIO transport
Handle all database failures gracefully
Use environment variables for credentials
Implement rate limiting for expensive queries
Add monitoring and alerting
Start Building
You now have a working MCP server that connects AI to PostgreSQL. The same pattern works for any external system: APIs, internal tools, file systems, cloud services.
MCP is now the standard. Build once, works everywhere.
Next steps:
Browse existing servers at the MCP registry
Build a server for your most-used internal tool
Contribute your server back to the ecosystem
Build your first MCP server this week and reply with what you built.
Also before you go, here’s a shoutout to today’s sponsors:
The Future of Shopping? AI + Actual Humans.
AI has changed how consumers shop, but people still drive decisions. Levanta’s research shows affiliate and creator content continues to influence conversions, plus it now shapes the product recommendations AI delivers. Affiliate marketing isn’t being replaced by AI, it’s being amplified.
Invest in Renewable Energy Projects Across America
Across America, communities are being powered thanks to investors on Climatize who have committed to a brighter future.
Climatize lists vetted renewable energy investment offerings in different states.
As of November 2025, over $13.2 million has been invested across 28 projects on the platform, and over $3.6 million has already been returned to our growing community of thousands of members. Returns aren’t guaranteed, and past performance does not predict future results.
On Climatize, you can explore vetted clean energy offerings, including past projects like solar farms in Tennessee, grid-scale battery storage units in New York, and EV chargers in California. Each offering is reviewed for transparency and provides a clear view of how clean energy takes shape.
Investors can access clean energy projects from $10 through Climatize. Through Climatize, you can see and hear about the end impact of your money in our POWERED by Climatize stories.
Climatize is an SEC-registered & FINRA member funding portal. Crowdfunding carries risk, including loss.
Until next time,
Vaibhav 🤝🏻
