---
title: "What is the Model Context Protocol (MCP)?"
excerpt: "An open standard for exposing tools and data to LLMs over one interface. It turns one-off integrations into reusable servers — when the tools are designed well."
definition: "The Model Context Protocol (MCP) is an open standard that defines how LLM applications connect to external tools and data sources through a uniform interface, so a tool built once can be reused across any MCP-compatible client."
date: "2026-07-09"
lastModified: "2026-07-09"
author: "Teo Deleanu"
authorAvatar: "/team/teo.jpg"
tags: ["MCP", "Agents", "AI Engineering", "Tool Calling"]
keywords:
  - "what is model context protocol"
  - "mcp explained"
  - "mcp tool server"
---

The Model Context Protocol (MCP) is a standard interface between an LLM application and the outside world — the tools it can call and the data it can read. Instead of hand-wiring every integration into every app, you build an MCP server once and any MCP-compatible client can use it. It is a common plug for [tool calling](/glossary/tool-calling).

## Why it matters in production

Before a standard, every model-to-tool connection was bespoke: new glue code per app, per model, per tool. MCP turns that into reusable servers with a defined contract. The value is not novelty — the underlying capability is still tool calling — it is that the integration becomes portable and testable instead of a tangle of one-offs.

## The common mistake

Assuming the protocol fixes tool design. MCP standardizes *how* a tool is exposed, not *whether the model can use it well*. A server that dumps forty overlapping, vaguely-described tools is just as confusing to a model as forty raw API calls — now with extra ceremony. The hard part, [designing a tool catalog](/blog/context-engineering-tool-catalog) the model can reason about, does not go away because you wrapped it in a standard.

## How we actually use it

We build MCP servers with a small, sharply-described tool set, treat each tool's schema as the model-facing contract it really is, and put [guardrails](/glossary/guardrails) on anything with side effects. We wrote up the concrete build in [inside an MCP tool server](/blog/inside-an-mcp-tool-server), and where MCP fits against plain retrieval and full agents in [MCP vs RAG vs agents](/blog/mcp-vs-rag-vs-agents).