Anthropic recently released the Model Context Protocol (MCP), an open standard that describes a protocol for integrating external resources and tools with LLM apps. This release includes an SDK that implements the protocol and an open source repository of reference implementations of MCP.
MCP aims to solve the “MxN” problem, the combinatorial difficulty of integrating M different LLMs with N different tools. Instead, MCP provides a standard protocol that LLM vendors and tool builders can follow. MCP uses a client/server architecture. AI apps such as Claude for Desktop and IDEs use MCP clients to connect to MCP servers that front data sources and tools. For developers who want to get started with MCP quickly, we have both Python and TypeScript SDKs, as well as a growing list of reference implementations and community-contributed servers. According to Antropic,
We are committed to building MCP as a collaborative open source project and ecosystem, and we welcome your feedback. Whether you’re an AI tool developer, a business looking to leverage existing data, or an early adopter exploring the frontier, let’s build the future of context-aware AI together. We recommend that you do so.
The MCP specification defines a set of JSON-RPC messages for communication between clients and servers. These messages implement constructs called primitives. The server supports three basic elements: prompts, resources, and tools. The client supports two things: routes and sampling.
The Server primitive is for “adding context to the language model.” A prompt is an instruction or a template for an instruction. Resources are structured data that can be included in an LLM prompt context. Tools are “executable functions” that LLM can call to obtain information or perform actions.
The root is the entry point into the file system and gives the server access to files on the client side. Sampling allows the server to request “completion” or “generation” from the client-side LLM. Anthropic says that you can use sampling to implement agent behavior by nesting LLM calls within server actions, but “there is always a human in the loop with the ability to reject sampling requests.” “There is a need,” he warns.
The documentation provides several examples and tutorials to show developers what they can build with MCP. The quickstart example shows how to use Claude LLM to get weather forecasts and alerts. To do this, the developer writes a Python program that implements the MCP server. This program exposes tooling primitives that wrap calls to public web services that return weather data. Developers can use LLM to call the MCP server and retrieve weather data via the desktop Claude app, which has a built-in MCP client.
Anthropic developer Justin Spahr-Summers joined the Hacker News discussion about MCP. When several users wondered if MCP could help solve the “MxN” problem, Spahr-Summers said, “We definitely hope so.” When asked how MCP differs from the use of existing tools in LLM, he replied:
Regarding tools in particular, there was some back and forth about whether other primitives in MCP were ultimately limited to tool usage, but ultimately the separate concepts of “prompt” and “resource” , I came to the conclusion that it is very useful for expressing different “intents” of server functionality. They all have a role to play!
The model context protocol specifications, documentation, and SDKs for Python and TypeScript are available on GitHub.
