The most important line in an MCP server is not code. It is the one plain-English sentence describing what your tool does, because that sentence is most of what the model ever sees.

MCP (Model Context Protocol) is an open standard for how AI apps talk to outside tools and data, released by Anthropic in 2024 and now stewarded by the Agentic AI Foundation under the Linux Foundation. An MCP server is the small program you write on your side of that line: it publishes a list of things a model is allowed to do, and does them when asked. Write it once and anything that speaks MCP can plug in: Claude, VS Code, your own app.

One tool is three parts. A name (get_order_status), a description, and a schema (a strict list of the inputs it accepts and their types). The body is boring code. Query a database, call an API, read a file. No AI runs inside it.

So the model never sees your implementation. It picks from the name, description and schema alone. Write a vague description and your tool sits unused, or fires at the wrong moment, and neither failure looks like a bug in your code.

That cuts both ways. Descriptions on someone else's server are text your model reads and acts on. Researchers showed one with instructions hidden inside it talking an assistant into forwarding a user's whole WhatsApp history to an attacker. Connecting to an unknown MCP server is closer to installing an unreviewed browser extension than importing a library.

Quick check before you scroll: Your team has 3 AI apps (Claude, a custom agent, VS Code Copilot) and wants all 3 to query the same internal database. Without MCP, how many custom integrations do you need? With MCP, how many?

Full breakdown + the answer: frankduah.me/learnings/2026-09-05-building-your-first-mcp-server

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 #AI #LLM #AIAgents #MachineLearning