In this video I show you how to take an existing REST API and wrap it into an MCP (Model Context Protocol) server so AI agents and LLM-powered apps can consume it cleanly. I start by breaking down why traditional APIs are “documentation-driven” and “hard-coded” from the client side—fine for normal apps, but brittle for LLM tool usage because you end up manually describing endpoints, parameters, and response parsing in prompts. Then I explain MCP (from Anthropic) as the “USB standard” for tools: servers dynamically describe their tools/resources/prompts, and the interaction becomes stateful and conversational. After that, we do a full hands-on build in three steps. First, I generate a small FastAPI weather service (current temperature + 5-day forecast, limited cities) and test it quickly. Second, I convert that API into an MCP server using the FastMCP Python framework, exposing tools like get_temperature, get_forecast, and list_cities, plus a resource—each tool internally calls the REST API but returns LLM-friendly output. Finally, I build an MCP client (also with FastMCP) that connects to the server, discovers tools automatically, and invokes them end-to-end. The key takeaway: MCP makes tool discoverability and integration way more reliable, and it sets you up for real agentic workflows where the LLM chooses which tool to call and with what parameters.