Specification

The Model Context Protocol (MCP) specification defines an open protocol designed to facilitate integration between large language model (LLM) applications and external data sources or tools. By establishing a standardized approach, MCP enables developers to connect LLMs with the context and capabilities necessary for advanced AI workflows, such as AI-powered IDEs, enhanced chat interfaces, and custom AI-driven processes. 0 messages for communication.

MCP structures communication between three main roles: hosts (LLM applications initiating connections), clients (connectors within host applications), and servers (services providing context and capabilities). The protocol supports stateful connections and negotiates capabilities between servers and clients. Servers can offer resources (context and data), prompts (templated messages and workflows), and tools (functions executable by the AI model). Clients, in turn, may provide features such as sampling (enabling agentic behaviors and recursive LLM interactions), roots (server-initiated inquiries into URI or filesystem boundaries), and elicitation (requests for additional user information).

Additional utilities within MCP include configuration management, progress tracking, cancellation, error reporting, and logging. The protocol emphasizes security and trust, requiring explicit user consent for all data access, operations, and tool invocations. Key principles include maintaining user control over data sharing and actions, ensuring data privacy through access controls, and treating tools with caution due to their potential for arbitrary code execution. Hosts must obtain user consent before exposing data or invoking tools, and users should be able to review and authorize all activities. The protocol also limits server visibility into prompts and sampling operations, further supporting user privacy and control.

The specification provides implementation guidelines recommending robust consent and authorization flows, clear documentation of security implications, and adherence to security best practices. While MCP cannot enforce these principles at the protocol level, it strongly encourages implementors to consider privacy and security in all integrations. The protocol is inspired by the Language Server Protocol, aiming to standardize integration of context and tools across the ecosystem of AI applications.