Building an MCP Server

Building an MCP server means turning a capability you own — an API, a database, a filesystem, a business process — into tools, resources, and prompts that any MCP host can use. The official SDKs make this deliberately small: a working server is a script that declares its tools with names, descriptions, and JSON Schema input definitions, implements a handler for each, and connects a transport. Most real servers fit in one file.

This is the highest-leverage skill in the module. Function calling locks your integration inside one application; an MCP server makes it portable across Claude Desktop, Claude Code, Cursor, VS Code, and every agent framework that speaks the protocol — which in 2026 is effectively all of them. It’s also where quality shows: the model chooses tools by reading your descriptions, so precise naming, tight schemas, and helpful error messages directly improve how reliably agents use your server.

In practice you’ll use FastMCP in Python — decorate a function with @mcp.tool(), type-hint the arguments, return a value — or McpServer in TypeScript with registerTool() and a Zod schema. Run it over stdio during development, test interactively with the MCP Inspector (npx @modelcontextprotocol/inspector), then register it in Claude Desktop’s config to use it for real. When you outgrow your laptop, switch the transport to streamable HTTP and deploy it like any web service — the tool code doesn’t change.