Model Context Protocol (MCP) Explained: The New Standard for AI Tools?

Key Concepts

• Model Context Protocol (mCP): A standard for connecting AI assistants to systems where data lives.
• Tool Calls/Function Calls: Techniques to extend LLM capabilities beyond their training data cutoff.
• JSON RPC 2.0: A lightweight remote procedure call protocol encoded in JSON.
• Hosts: Applications running AI models (e.g., cursor, windsurf).
• Clients: Modules within hosts responsible for communication with servers.
• Servers: Lightweight programs exposing specific capabilities through the mCP standard.
• Resources: File-like data that can be read by the client (e.g., API responses, file contents).
• Tools: Functions that can be called to perform actions (e.g., sending an email, querying a database).
• Prompts: Templates for structuring interaction with the LLMs.
• Language Server Protocol (LSP): Standardized how to add support for programming languages across whole ecosystem of development tools.

What is Model Context Protocol (mCP)?

The Model Context Protocol (mCP) is an open-source protocol introduced by Anthropic as a new standard for connecting AI assistants to systems where data lives, including content repositories. It aims to standardize the interaction between tools and Large Language Models (LLMs). The protocol uses JSON RPC 2.0 messages to establish communication between three components: hosts, clients, and servers.

The Problem mCP Solves

LLMs have a limited knowledge base due to the cutoff date of their training data. Techniques like Retrieval-Augmented Generation (RAG), tool calls, and function calls extend their capabilities. However, these methods often require custom API integrations for each tool (e.g., calendar, email, database). This leads to two main issues:

1. Hallucination: As the number of tools increases, the LLM may struggle to decide which tool to use.
2. Lack of Standardization: Each tool requires a custom integration, and changes to the underlying API implementation necessitate updates to the tool integration.

How mCP Works: Components and Communication

mCP introduces a standardization layer between the AI application (host) and the actual tool implementation (server). This layer remains consistent even if the tool is updated, as long as it adheres to the mCP standard.

• Hosts: AI applications like cursor, Cloud, open AI, windsurf, or custom-built applications that need access to external data or tools.
• Clients: Modules within hosts that handle communication with servers, maintaining connections and forwarding requests.
• Servers: Lightweight programs exposing specific capabilities through the mCP standard. They connect to local data sources (files, databases) or remote services (external systems via APIs).

The host uses the client to communicate with the server, which then interacts with the external world through custom APIs or tools.

mCP vs. Traditional Tool Usage

In traditional tool usage, the LLM needs to track and manage each tool manually. This involves providing a list of available tools, manually describing them, and updating them whenever the API changes. The LLM decides which tool to use, makes the call, gets the response, and feeds the results to generate the final response.

mCP unifies this process by abstracting away the complexities of individual tool implementations. The LLM only needs to know which mCP server to use, not the specific tools available on that server.

Example:

To summarize the five latest commits from a repository:

1. The mCP host (AI application) and client call the mCP server to ask which tools are available.
2. The LLM receives the information and chooses to use a tool.
3. It sends a request to the mCP server via the host.
4. The server uses the tool and returns the results, which are fed back to the LLM.

Key Components of an mCP Server

Each mCP server exposes three main components:

1. Resources: File-like data that can be read by the client, such as API responses or file contents.
2. Tools: Functions that can be called to perform actions, such as sending an email or querying a database.
3. Prompts: Templates for structuring interaction with the LLMs, such as document question answering.

Why the Hype Around mCP?

According to Sean from Latent Space podcast, the reasons for mCP's increasing popularity are:

1. Becoming a Defacto Standard: It's an alternative approach to tool calls, with potential to become the standard for data connection between LLMs and data sources.
2. Open Standard with Big Backer: Anthropic is actively developing and updating mCP, and companies like open AI and Google are considering adopting it.
3. Anthropic's Developer AI Brand: Cloud is a well-regarded coding LLM.
4. Based on LSP: mCP is inspired by the successful Language Server Protocol (LSP), which standardized language support across development tools.

Community Adoption and Cautions

A community is forming around mCP, with tools like cursor and windsurf supporting mCP servers. However, it's crucial to vet mCP servers carefully, as they can expose API keys and data.

Conclusion

mCP is a promising standard for simplifying and standardizing the integration of external tools and data sources with LLMs. While its future is uncertain, it currently represents a leading approach to addressing the challenges of tool usage in AI applications. It's important to note that mCP may not be necessary for all use cases, especially those with a small number of tools.