What Is MCP?
Last updated: 2026-05-18
What Is MCP (Model Context Protocol)?
MCP (Model Context Protocol) is an open protocol from Anthropic that lets AI models connect to external tools and data sources. By default, models are isolated — they only see what you type. MCP gives them the ability to read files, query databases, call APIs, and use other tools. It's a standard way for AI applications to extend what they can do.
If you've used Claude in Cursor or another MCP-enabled client, you've seen it in action. The model can browse your codebase, run commands, or fetch data because MCP servers expose those capabilities as tools the model can call.
The Problem MCP Solves
Models are powerful at reasoning and generating text, but they have no direct access to:
- Your file system
- Databases
- External APIs and web services
- Real-time information
Before MCP, every application built its own integration layer. There was no common way to say "this model can use these tools." MCP defines that interface. Servers expose tools; clients connect to servers and let the model invoke them.
How MCP Works
Servers — Software that exposes tools and resources. A filesystem server might expose "read_file" and "list_directory." A database server might expose "query." Servers run locally or remotely and communicate over the protocol.
Clients — AI applications (Claude desktop, Cursor, etc.) that connect to MCP servers and pass tool calls to the model. The model decides when to call a tool; the client executes the call and returns the result.
Tools — Named actions the model can invoke. Each tool has a name, description, and parameters. The model sees the tool list and chooses which to call based on the request.
Resources — Read-only data the model can access (e.g., document contents, configuration). Different from tools, which perform actions.
Real-World Examples
- Filesystem — Model reads and writes files in a project directory
- Database — Model runs queries against Postgres, SQLite, or other DBs
- Web — Model fetches URLs, scrapes pages, or calls REST APIs
- Git — Model checks status, diffs, and branch info
- Slack/Discord — Model sends messages or reads channels
Developers and power users run multiple MCP servers so a single model session can access many tools at once.
MCP vs. Function Calling vs. Plugins
Function calling — A general term for models invoking external functions. OpenAI, Anthropic, and others support it in their APIs. MCP is one implementation: it standardizes how those functions are defined and discovered.
Plugins — App-specific extensions (e.g., ChatGPT plugins). Each platform has its own plugin system. MCP is platform-agnostic: the same server works with any MCP-compatible client.
MCP doesn't replace function calling; it provides a portable, open way to define and expose tools across clients.
The Ecosystem in 2026
MCP adoption is growing. Anthropic's Claude desktop and Cursor support it natively. Other IDEs and AI tools are adding support. The protocol is open, so anyone can build servers or clients. Popular servers cover filesystems, databases, web search, and productivity apps.
Tools with MCP support integrate more easily into a connected stack. A model that can call your database, read your docs, or trigger workflows is more useful than one that can't. When evaluating AI tools, check whether they support MCP or similar tool-calling standards.