MCP (Model Context Protocol)

Model Context Protocol (MCP) is an open standard introduced by Anthropic in late 2024 that defines a universal way for large language models to communicate with external systems — files, databases, APIs, and developer tools — without requiring custom integrations for every combination.

The Problem MCP Solves

Before MCP, every AI application had to build its own integrations: Cursor connected to your filesystem one way, Claude.ai another, and each tool had its own format for passing context. This created an M×N integration problem — M AI systems × N tools.

MCP reduces this to M+N: tool providers publish one MCP server, and AI clients implement one MCP client. The protocol handles discovery, transport, and structured message exchange.

How MCP Works

MCP uses a client-server model over standard I/O or HTTP with Server-Sent Events:

• MCP Server — exposes resources (files, DB rows), tools (functions the AI can call), and prompts (reusable prompt templates).
• MCP Client — the AI host application (Claude Desktop, Cursor, OpenHands) that connects to one or more servers.
• Transport — stdio for local servers; SSE for remote servers.

When a user asks Claude to "check my database for slow queries," Claude uses the MCP client to call the database server's query tool, receives results in a structured JSON format, and incorporates them into its response — all without the developer writing custom glue code.

Core Primitives

• Resources — read-only data (files, API responses, git history)
• Tools — executable functions (run SQL, call REST API, write file)
• Prompts — parameterized prompt templates for common tasks
• Sampling — servers can request LLM completions through the client

MCP in Practice

By early 2025, the MCP ecosystem had hundreds of servers: GitHub, Slack, PostgreSQL, filesystem, web fetch, and more. AI editors like Cursor and Windsurf ship with MCP support. Openhands uses MCP to give agents access to dev tools. Companies building on Langchain or direct API calls are adding MCP servers to expose their internal tools to LLMs.

Why It Matters

MCP is becoming the de facto standard for agentic AI. As models become more capable of multi-step reasoning and tool use, a reliable, secure protocol for tool invocation is foundational. The security model (user approval for tool calls, capability-scoped permissions) also addresses a key concern in Prompt Engineering: preventing prompt injection attacks that trick the AI into calling unintended tools.

Related Terms

• Langchain — popular framework with its own tool-calling abstractions; MCP is an alternative protocol
• Prompt Engineering — how you design the instructions that guide an LLM to use MCP tools correctly
• Openhands — open-source agent that uses MCP for tool access
• Cursor — AI code editor with native MCP support