Protocols & Standards

MCP

Also called: Model Context Protocol

MCP (Model Context Protocol) is an open standard that lets AI applications connect to external tools, data, and prompts through one consistent interface instead of a custom integration for each.

Host / Client appMCP ServerMCP · JSON-RPC 2.0
A host application's MCP client connects to an MCP server over JSON-RPC 2.0 to call tools, read resources, and fetch prompts.

Before MCP, every AI application that wanted to give a model access to a tool — a database, a file system, a search API — had to write its own custom integration code for that specific combination of app and tool. MCP standardizes that connection so a tool only needs to be wired up once, as an MCP server, and any MCP-compatible application can use it.

MCP was originated by Anthropic and released as an open standard, with its specification published and since moved toward broader, community-driven governance. It has become the most widely referenced open standard for connecting AI applications to external context and capabilities, though the space is still actively evolving and other approaches exist alongside it.

In MCP's architecture, a host application (like a coding agent or chat client) runs an MCP client, which opens a connection to one or more MCP servers. Each server can expose three kinds of primitives: tools the model can call, resources it can read for context, and prompt templates it can reuse. The host decides what the model is allowed to see and invoke.

How it works

MCP communication runs over JSON-RPC 2.0, typically over stdio for local servers or HTTP-based transports for remote ones. When a client connects, it asks the server what it supports — a capability negotiation step — and the server responds with its available tools, resources, and prompts. From then on, the model (via the host) can call a tool with structured arguments, request a resource by its URI, or fetch a prompt template, and the server returns a structured result. Because the protocol is the same regardless of what the server actually does behind the scenes, a single MCP client implementation can talk to a filesystem server, a database server, and a project-management server without any server-specific code in the host application.

Example

A coding agent with an MCP client might connect to a github MCP server to open issues, a postgres MCP server to query a database, and a filesystem MCP server to read local files — all through the same protocol, with the model deciding at runtime which tool to call for a given step.

How it differs

MCP is sometimes confused with plain function/tool calling. Function calling is the underlying mechanism a model uses to invoke a structured call — MCP standardizes how that call, along with resources and prompts, gets exposed and discovered across many different tools and applications, so the model doesn't need a bespoke integration for each one.

Common misconceptions

Often assumed: MCP is an Anthropic-only or Claude-only feature.
Actually: MCP is an open, published specification; multiple AI applications and vendors beyond Anthropic have built MCP clients and servers.
Often assumed: MCP replaces function calling.
Actually: MCP doesn't replace function/tool calling — tools exposed by an MCP server are still invoked the same way; MCP standardizes discovery and connection, not the underlying call mechanism.

FAQ

What is MCP used for?
It's used to connect AI applications — coding agents, chat assistants, IDE plugins — to external tools, data sources, and reusable prompts without writing a custom integration for each combination.
Is MCP only for Claude?
No. MCP is an open standard; it originated at Anthropic but its specification is public, and other AI applications and tool vendors have built their own MCP clients and servers.
Do I need to build my own MCP server?
Only if you're exposing a new tool or data source that doesn't already have one; many common tools already have community- or vendor-built MCP servers available.

Last checked: 2026-08-28

Related terms