Ten AI apps that each need to reach ten services used to mean up to a hundred separate pieces of custom glue code.

MCP (Model Context Protocol) cuts that to about twenty. The official docs compare it to USB-C: one shared plug instead of a different cable for every device.

Here is how it works. Each service gets wrapped once in an MCP server, a small program that publishes what it can offer the model: tools (actions it can run, like "create a ticket"), resources (data it can read) and prompts (reusable templates). Each app, such as Claude Code or Cursor, learns to speak MCP once. Then the app asks a server which tools it has with one request, tools/list, and runs one with another, tools/call.

The part I find most useful is how failures come back, because a model working through a task on its own has to read the result and pick its next step.

A broken request, like a tool name that doesn't exist, is a protocol error: rejected before anything runs. A tool that ran but hit a problem, like a service refusing because too many requests came in, returns a normal result marked isError: true with a plain message the model reads. "Rate limit exceeded, retry in 60 seconds" gives it something to act on. A raw crash dump usually doesn't.

One habit worth keeping: connect only the servers a task needs. Every tool listed is something the model has to read and choose between, so more is not better.

Quick check before you scroll: An MCP server lists a tool called weather_current. What two requests does a client use to find it and then run it?

Full breakdown + the answer: frankduah.me/learnings/2026-10-11-mcp-in-plain-english-usb-c-for-ai-tools

New here? I post a bite-size AI / ML concept like this every day. Follow me for the daily drop, and it compounds fast. Why I do it: https://lnkd.in/gK8knHDH

#MCP #ModelContextProtocol #AgenticAI #AI #LLM #AIAgents #MachineLearning