Model Context Protocol

Model Context Protocol
Developer	Anthropic
Status	Stable
First published	November 2024
Transport	HTTP, stdio
Encoding	JSON-RPC 2.0
License	MIT
Website	https://modelcontextprotocol.io
Specification	https://spec.modelcontextprotocol.io
Repository	https://github.com/modelcontextprotocol

Main article: wp:Model_Context_Protocol

The Model Context Protocol (MCP) is an open standard developed by Anthropic that provides a standardized way for large language models to connect with external tools, data sources, and APIs. Often described as the "USB-C port for AI," MCP enables AI applications to access real-time information and perform actions beyond their training data.

MCP was created by David Soria Parra and Justin Spahr-Summers at Anthropic and deliberately reuses message-flow concepts from the Language Server Protocol (LSP). Following its release, MCP was adopted by major AI providers including OpenAI (March 2025) and Google DeepMind (April 2025).

Overview

MCP addresses a fundamental challenge in AI application development: integrating language models with external systems. Before MCP, developers needed to create custom integrations for each tool and data source, resulting in fragmented and difficult-to-maintain codebases.

Design philosophy

MCP follows several key principles:

• Vendor-agnostic – Works with any LLM provider, not just Anthropic
• Standardized interfaces – Consistent patterns for tool definition and invocation
• Security-focused – Built-in access control and audit capabilities
• Human-in-the-loop – Support for requiring human confirmation of actions

Architecture

MCP uses a client-server model with three main components:

MCP Host

The host contains orchestration logic and manages connections between clients and servers. It can host multiple clients simultaneously.

MCP Client

Clients convert requests into the structured MCP format. Each client maintains a one-to-one relationship with an MCP server and handles session management, response parsing, and error handling.

MCP Server

Servers expose tools, data sources, and APIs to clients. They are typically implemented as standalone applications or services that provide access to specific capabilities.

Technical specification

Transport layer

MCP supports two transport mechanisms:

• stdio – Standard input/output for lightweight, synchronous messaging
• SSE over HTTP – Server-Sent Events for asynchronous, event-driven communication

Message format

All messages use JSON-RPC 2.0 format:

Example tool call

{{{code}}}

Core primitives

MCP defines four core primitives:

Primitive	Description
Resources	Data sources and files accessible to the model
Tools	Functions the model can invoke
Prompts	Pre-defined prompt templates
Sampling	Model inference requests

Adoption

Since its launch in November 2024, MCP has seen rapid adoption:

• Integrated into Visual Studio Code via GitHub Copilot
• Supported by AWS for agentic AI development
• Hundreds of community-developed MCP servers available

Comparison with other protocols

MCP focuses on tool integration ("vertical" connections), while protocols like A2A handle agent-to-agent communication ("horizontal" connections). The two are complementary rather than competing standards.

See also

• Agent-to-Agent Protocol
• Agent Communication Protocol
• Common Text-based Response Protocol
• Tool Calling

External links

• Official website
• Protocol specification
• GitHub organization