Model Context Protocol

The Model Context Protocol (MCP) is an open standard for connecting large language model (LLM) applications to external data sources and tools. It standardizes how AI systems share context, expose capabilities, and compose integrations. ^[1]

Architecture

MCP uses a client-server-host architecture to separate concerns:

• Hosts: LLM applications (such as Claude Desktop, IDEs, or custom agents) that initiate connections.
• Clients: Connectors within the host that maintain the protocol session.
• Servers: Lightweight programs that provide specific capabilities like data access or tool execution. ^[2]

Communication is handled via JSON-RPC 2.0 messages, allowing for a structured and language-agnostic interface. ^[1]

Transports

MCP supports two primary transport mechanisms for message exchange:

• stdio: The client launches the server as a local subprocess and communicates via standard input/output. This is common for local tools and desktop integrations.
• Streamable HTTP: The server runs as a separate network service. It uses HTTP POST for client-to-server messages and optionally Server-Sent Events (SSE) for server-to-client notifications. ^[3]

Capabilities

Servers can register three main types of capabilities:

• Resources: Read-only data sources (like files, database records, or API responses) that provide context to the model.
• Tools: Executable functions that allow the model to perform actions in the external world. ^[1]
• Prompts: Reusable templates that allow servers to define common workflows and standard interactions.
  • Structure: A prompt is defined by a unique name, a description, and a set of arguments. It returns a list of messages (such as user or assistant roles) that can be fed into an LLM.
  • Usage: Prompts are typically surface in the host application's UI as slash commands (e.g., /debug-error) or menu items. When selected, the client renders the prompt—potentially asking the user to fill in arguments—and then inserts the resulting context and instructions into the conversation. This allows servers to provide "expert" guidance on how to best use their resources and tools. ^[4]

Sampling

A distinctive feature of MCP is Sampling, which allows a server to request an LLM completion back from the client. This enables agentic behaviors where a server-side tool can "ask" the model to analyze data or make decisions as part of its execution. Crucially, the client retains control over which model is used and can require user approval for any sampling request, ensuring that the human operator remains in the loop. ^[5]

History and Adoption

MCP was announced and open-sourced by Anthropic in November 2024. ^[6] Since then, it has seen growing adoption across the AI ecosystem, including integration into development environments like Cursor and localized toolchains for specialized data analysis. ^[7]