MCP protocol Specification

The Model Context Protocol (MCP) is an open protocol enabling seamless integration between LLM applications and external data sources. This specification defines the authoritative protocol requirements for building AI-powered integrations.

MCP specification

MCP uses JSON-RPC 2.0 messages with a client-host-server architecture. The protocol provides standardized ways to share contextual information, expose tools, and build composable integrations with security and trust considerations.

MCP architecture

The MCP architecture follows a client-host-server model where hosts manage multiple client instances. Each client maintains a stateful session with a server. Key design principles include making servers easy to build, highly composable, and maintaining security boundaries between servers.

MCP base protocol

The base protocol covers core MCP components including JSON-RPC messages, connection lifecycle with three phases, and standard transports (stdio and streamable HTTP). It also addresses authorization using OAuth 2.1 and optional utilities like request cancellation and progress notifications.

MCP client features

Client features include roots for exposing filesystem boundaries, sampling for LLM generation requests, and elicitation for gathering user information. These features enable controlled access, tool usage with user approval, and standardized information gathering with privacy considerations.

MCP server features overview

This page provides an overview of MCP server features, covering resources, prompts, and tools that servers can expose to clients. Servers operate independently with focused responsibilities and can be local processes or remote services.

MCP schema reference

The schema reference provides a comprehensive guide to MCP JSON-RPC types, common types, content types, and protocol messages. It serves as the definitive reference for implementing MCP-compliant applications.

Key changes in MCP

This document lists all changes since the previous revision (2025-06-18). Major updates include enhanced authorization with OpenID Connect, tool icons support, incremental scope consent, URL mode elicitation, and experimental task tracking support.