What Is MCP: How It Works and Why It Matters

Key takeaways

• MCP is a communication layer, an open-source standard that enables AI agents and LLMs to communicate with external tools and applications.

• Developers like MCP because it allows AI to negotiate with tools and applications to perform the desired actions without requiring complex coding, while users enjoy the more accurate, real-time results and expanded functionality it provides.

• MCP works by verifying compatibility between the AI agent or LLM and the external application or tool, communicating the desired action between the MCP host, client, and server, and receiving the results before finalizing the requested task.

Learn more about what MCP is and why it matters in AI, including details about its architecture and benefits. To begin building your AI skills and learning about implementing AI, enroll in the AI For Business Specialization from the University of Pennsylvania. In as few as four weeks, you'll have an opportunity to learn the fundamentals, including AI ethics, how to design frameworks for fair use, and how to deploy AI for optimal results.

What is MCP?

Artificial intelligence (AI) research company Anthropic, which also released the AI tool Claude in April 2022, created the Model Context Protocol (MCP) as a standard for secure, open-source, two-way connections between AI applications, data sources, and other external tools, such as calculators, workflows, and search engines [1, 2].

In doing so, this open-source framework solves two challenges you might encounter when working with large language models and AI agents: accessing real-time data and interacting with other agents or outside applications. Using MCP enables AI to take on a more dynamic role, with expanded capabilities, enhanced accuracy, and more context-aware results. It impacts everyone from developers to end users, shortening development time, enhancing the broader AI ecosystem, and providing a better user experience.

For example, you could connect ChatGPT or an AI agent, which would otherwise be limited to its training data, to your customer relationship management (CRM) software or sales system’s MCP and then tell it to search that system for a specific transaction or create a report. The MCP connection enables your AI tool to do more for you, with access to real-time information and a greater ability to connect a multi-agent collaborative ecosystem.

What is an MCP vs. API?

An application programming interface (API) is similar to an MCP because it provides protocols and a framework for applications to communicate. However, the two connect in different ways. While an API provides rules for one software to communicate with another, it is a hard-coded connection that requires developers to anticipate in advance. MCP allows AI to negotiate with an MCP server, enabling it to determine how to perform the functions users request without further developer involvement [3].

Both offer benefits and unique use cases. To illustrate when to use one versus the other, consider the following:

• Use an API: If you integrate two systems with predictable interactions, using an API can enhance the speed and precision of the connection while also making debugging easier. For example, you can use an API to connect a CRM with financial reporting because it requires minimal connections and a greater degree of control over the requests.

• Use an MCP: When connecting multiple AI-powered integrations with multiple tools, an MCP streamlines the process. For example, with an API, integrating four AI agents with five tools would require 20 individually coded integrations. MCP creates a universal language that lets those agents communicate with the tools and connect accordingly.

What is the difference between MCP and LLM?

An LLM is a type of AI system, while MCP is a standard that enables LLMs to communicate with other tools. LLMs, such as ChatGPT, Google Gemini, or Anthropic Claude, provide results in human-like language. They train on immense datasets and use deep learning techniques to perform various tasks, including writing content, solving complex problems, and answering questions [4].

Learn more: How Do Large Language Models Work? How AI Understands and Generates Text

MCP architecture

You might hear people refer to the MCP architecture as a client-host-server model because it typically consists of an MCP server, client, and host. Each component plays a role in the MCP workflow. Explore each component in more detail, along with the two layers that support its functionality [5].

MCP host

The MCP host acts as a coordinator, providing centralized infrastructure to manage MCP client and server communication. This component contains the AI agent or LLM and is the part of MCP that users interact with. For example, when the LLM needs to communicate with an external application, the host launches and connects to the application's MCP server and client.

MCP client

This component initiates access to external resources, translating the LLM's request to the MCP server and then passing the MCP server's response back to the LLM. Each client connects to a single server and handles the communications between the host and the server.

MCP server

An MCP server is external to the LLM or agent and provides the AI with data, tools, context, or capabilities. MCP servers convert the AI's user request into an action the server can take, such as retrieving data from databases, connecting to other tools for information, and executing the user's commands through prompts. Various types of servers provide different integrations. For example:

• Local data sources: Servers that connect to your computer's information, including files and local databases

• Remote services: Connect to external tools like cloud-based databases

• Community servers: Developer-built and shared servers

• Official integrations: Organizations offer these servers to connect with services (for example, the Slack MCP server is Slack’s official integration)

Layers

In addition to the three main components that make up the MCP architecture, it also contains two layers: data and transport. The data layer enables client-server communication. The transport layer defines the channels and mechanisms by which data moves between servers and clients.

How it works: MCP in action

Understanding MCP architecture is only part of the equation. The initial greeting, sometimes called a handshake protocol or capability discovery, verifies that both the MCP server and client can communicate. After establishing compatibility, it can begin the process.

Explore a simplified version of how MCP works:

• A user opens an AI-powered, MCP-enabled tool and asks the LLM to perform an action.

• The LLM activates the MCP client to search for relevant tools before generating a request to use them. The MCP client sends the request to the tool's MCP server.

• Once the MCP server receives the request, it performs the action, gathers the information, and formats it to send to the LLM.

• The LLM receives the information and performs the desired action as prompted by the user, then confirms that it has completed the request.

What would this look like in the real world? Imagine you want to pull sales figures from your company’s database, create a visualization, and email it to the sales team. You might experience the following:

1. You ask your AI agent to gather sales data from the first quarter in the XYZ database and create a visualization using your preferred program.

2. The LLM will handle the handshake protocol and send the request to the database's MCP server.

3. The server will retrieve the information and format the data into a report to send to the LLM.

4. The LLM will receive the data and perform the second action, contacting the visualization software's MCP server to create the visualization.

5. The visualization software's MCP server will generate the report and send it back to the LLM.

6. The LLM will connect to the email sending tool and provide your email address and the contents of the report before sending it to the sales team.

7. The LLM will send you a confirmation that it has completed all steps.

Benefits and limitations of MCP

MCP enables AI to work outside of its training, transforming it from a tool that relies on static training into one that works in more dynamic environments. It simplifies connecting AI to applications, eliminating the need for complex custom integrations.

• Enhance AI usability: By enabling AI to interact with your chosen applications, tools, and databases, you can prompt your AI to perform tasks rather than generate responses.

• Increase automation potential: MCP allows AI to run an ongoing loop while interacting with multiple external tools, simplifying complex workflows and automating across multiple data sources and systems.

• Universal connector: MCP acts as a middle layer between AI and other tools, converting input and output in a manner the model can understand.

• Lighter development loads: With APIs, engineers have to manually code connections, manage permissions, and perform debugging, but MCP makes it easier to connect and switch between various implementations and AI providers.

• Centralized control: MCP gives you one point of control where you can manage all permissions, updates, and access protocols across AI clients, making it easier to maintain compliance, conduct audits, and ensure security protocols remain enforced.

• Fewer hallucinations: LLMs can provide incorrect information or make up facts in their responses. MCP reduces these hallucinations by giving AI a path to access reliable sources for more accurate, true responses.

Potential limitations

MCP implementation offers several advantages. However, it also comes with considerations to keep in mind. For example, it's still evolving and maturing, and it's a standard, not a platform. Organizations building on top of MCP must take responsibility for providing security frameworks, governance, data quality, and tool reliability. MCP relies on AI tools to perform the work, which introduces potential security risks. When working correctly, MCP servers execute an action with a user's permission, but that isn't guaranteed. Instead, a misconfigured MCP server implementation can affect authorization and authentication. Potential concerns include:

• User consent and control

• Secure handling of LLM outputs

• Safety of the tool integrated with the LLM

• Data privacy concerns

• Potential for bias or breaches in the chain between external tools and MCP servers

MCP vs. RAG: What’s the difference?

Retrieval-Augmented Generation (RAG) is another method LLMs can use to access real-time data from external sources. The main goal of RAG is to help the LLM access up-to-date, relevant information before it responds to a user's prompt. MCP, on the other hand, standardizes communication between LLMs and external services, data sources, and tools, enabling the LLM to not only retrieve information but also perform prescribed actions. RAG could be the preferred approach when working with a large body of knowledge that changes relatively infrequently and includes unstructured documents such as blogs, articles, and PDFs [5].

MCP use cases

Initially, MCP established a standard among AI engineers and facilitated development. Today, its adoption goes beyond developers and engineers and encompasses knowledge workers and well-known companies. For example, the payment provider Stripe and the web builder Wix have both implemented MCP [6].

Additionally, thousands of publicly available MCP servers support a wide range of uses, including coding, checking the weather, finding flights, and analyzing medical records. Some of the primary use cases fall into one of four categories: decision support systems, e-commerce, context-aware process automation, and data integration and knowledge management.

How to choose an MCP server

Each MCP server caters to different uses, so your primary objectives should be among the top considerations. When choosing an MCP server, also consider scalability, performance requirements, and security. You might also consider MCP server certification, which can help ensure reliability and security compliance.

Explore our free AI resources

Get fresh insights every week about emerging technologies, career paths, and in-demand skills with a subscription to Career Chat, our LinkedIn newsletter. Also, learn more about AI with our other resources, such as:

• Learn from an expert: The AI Advantage: 9 Questions with UC Davis AI Instructor Sadie St. Lawrence

• Watch on YouTube: How Agentic AI Works: the Flipped Interaction Pattern

• Find your path: AI Learning Roadmap: From Beginner to Expert

You can also develop your skills in AI and other key technologies with Coursera Plus. You can learn and earn credentials at your own pace from over 350 leading companies and universities. With a monthly or annual subscription, you’ll gain access to more than 10,000 courses.