Skip to main content

Is MCP the TCP/IP for AI?

 

The Internet runs on TCP/IP. We might as well include AI in the mix, but how about AI? TCP/IP allows devices, services, applications, and data to communicate with other devices, services, applications, and data. How does AI talk? This is why I love networking, and the thought of how AI communicates makes me hopeful that the field will survive extensive job replacements. Unless self-healing networks actually come. 

I have been reading more into MCP. What exactly is it? Is it necessary for AI models and agents to communicate? Did you know AI agents have their own social media?? Biggest of all, is there any correlation between MCP and TCP/IP for model communication? Can MCP take notes from TCP/IP, or will it rewrite networking?

What is MCP?

To understand MCP, a good place to start is APIs. To understand APIs, the example often used is a diner at a restaurant. When a person orders their food, they do not go straight into the kitchen and get their food. Their waiter or electronic kiosk nowadays takes the order, sends it to the pass for the chefs to make it (another computer with an electronic kiosk), and returns it to the diner. 

The thing is not every restaurant has the same way for their customers to order food. Some rely on kiosks with a register to help those who do not want to use it. Others have fully-manned registers. In addition, if you use food services like DoorDash or Instacart, you can make mobile orders. Similarly, different companies have different API integrations, which can make it harder for services to communicate if they do not follow a similar structure. MCP evens the playing field.

The Model Context Protocol is a structured, consistent way for AI models to request services, data, tools, and context from external sources. It is standardized, which can reduce custom integrations, maintain system consistency, and manage compatibility issues. 

Is it Necessary?

Simply, no. Why?
  • MCP is one way AI models communicate and access tools, data, and context. We often use GenAI tools, such as Gemini, ChatGPT, and Claude. We use human language in our prompts, but it is translated into requests so the system knows what tools and resources to access. 
  • MCP was created by Anthropic in late 2024. In the IT field, it is relatively new. 

Why Relate MCP to TCP/IP?

For one thing, the last letter in the acronym is P, which stands for Protocol. The TCP/IP is the backbone of the Internet. The Internet is built upon protocols. Where am I going with this? 

It is not perfect nor is it inherently secure, but many services rely on open-source protocols, such as HTTPS, for communication between internal and external systems and resources. In addition, many architectures follow a client-server model. While the TCP/IP model helps explain how devices within a network communicate, MCP accomplishes a similar purpose between AI models and systems. It is designed as an open source protocol that allows AI models to request resources exposed by systems. This is important because information may not be publicly available.

For example, I have done years of certification preparation. To organize my notes and have a custom knowledge base, I am looking to build an LLM to query my notes and make calls to expand on the topic, generate ideas for labs, and refresh my memory. My notes are not publicly accessible to an AI system nor to the public. However, I can use a tool that will allow the AI model running my LLM to access my notes.

This architecture shows how MCP connects an LLM to notes, tools, and data sources while RAG provides the context used to generate responses.

Why does this matter? 

Should it? I think there are lessons in how TCP/IP works that can help us understand how MCP works.

  1. Standardization beats customization.

    In a world of proprietary networking standards, such as AppleTalk and IPX, TCP/IP remains king, as it is how applications are able to communicate across hosts. Anthropic boasts that MCP "provides a universal, open standard for connecting AI systems with data sources, replacing fragmented integrations with a single protocol."

  2. Balance the playing field by implementing interoperability.

    This supports the previous point. The TCP/IP stack allows for communication across devices because of how it is designed. MCP, as an open standard, "enables developers to build secure, two-way connections between their data sources and AI-powered tools". 

  3. Understand layered communication.

    An important part of networking is understanding how data turns into bits and vice versa. That is why there are two main models that help us visualize how networking works: OSI and TCP/IP. AI uses models to ingest input, reason, and provide output. MCP, albeit a protocol at layer 7, is the middleman between LLMs and AI systems. Understanding its role as a layer to request and transfer information provides clarity to an abstracted process.

  4. Don't forget protocols.

    Protocols are agreed rules for communication. Whether it is host-to-host or full encapsulation-decapsulation, systems can communicate because of how protocols consistently operate. For AI, that can be hard due to its non-deterministic nature. How can we bring consistency to something unpredictable? MCP follows a client-server model, where the AI model is the client and calls to a server with tools, integrations, or API hooks. The model queries the server, the server returns results, and those results inform the final response delivered to the user. Protocols provide reliability around non-determinism. 

  5. Apply AAA to MCP security.

    This introduces significant security considerations within MCP. You are allowing an open protocol to access local files and tools. If not properly managed or allowed malicious or incorrectly worded prompts, this poses a major security risk, calling for robust authentication, authorization, and auditing. Where should it be implemented: at the client, the server, the LLM? Fact of the matter, it should be everywhere. If we can standardize MCP, there should be a MCP security counterpart.

These parallels become clearer when viewed side by side. 

MCP and TCP/IP Parallels
Lesson TCP/IP Parallel MCP Parallel Significance
Standardization beats customization TCP/IP as the backbone and universal communication standard of the Internet MCP provides a universal way for models to request tools and data Creates a template for AI systems where models can plug into many services and tools without custom integration
Interoperability enables cross-platform AI ecosystems Provided a practical standard for all devices, not just a class of devices Different models and tools connect through a shared protocol Open standards allow AI systems to integrate across platforms, reducing vendor lock-in and enabling ecosystem growth
Layered communication simplifies complexity TCP/IP model uses 4-5 layers to explain how data travels across a network MCP serves as a layer between client requests and data, context, and tool retrieval This allows for scalable AI systems and structured understanding of troubleshooting and debugging issues
Protocols provide a foundation Client-server communication patterns MCP as a protocol follows a client-server architecture: MCP clients interact with MCP servers Brings consistency to non-determinism
AAA TCP/IP is inherently insecure, requiring layered security controls such as authentication, encryption, and monitoring MCP architecture must consider identity, permissions, and auditing to govern tool access Ensures AI systems access tools and data securely, with accountability and traceability
 
It doesn't have to be confusing. I am still learning it, reading documentation. The next step is building a small learning environment where a model can interact with external tools and data sources and my notes, turning theorizing MCP into understanding MCP. Understanding MCP conceptually is one step. Building with it is the next. How about you? How do you feel about how MCP is implemented within companies and projects?

Comments

Post a Comment

Popular posts from this blog

IP in Practice: The Need for IPv6

 It is time for IPv6. Not just for our network infrastructure, but for this series. The IPv4 section has come to an end, and now it is time for a deep dive into IPv6. More than ever, it is time to consider systems that can handle heavier workloads, more devices, and fewer address limitations - AI agents, IoT devices, edge computing. This post will examine why we need IPv6 and why it is an important network solution.   Why do we need IPv6? We need more space. IPv6 literally increases the address space exponentially. The IPv6 address space is 2 128 total addresses, 3.4 followed by 38 zeroes. Technology is no longer limited to servers, office computers, and mobile devices. Today's systems integrate AI infrastructure, edge computing, IoT devices, cloud networking, virtual machines, and more. Simply, more devices mean more space.  We need more scalability. As more devices connect to systems, administrators need to consider not only the number but the distribution. System...
1 comment
Read more

IP in Practice: IPv4 Address Structure & Classes

There are two main types of IP addresses: IPv4 and IPv6. Many of us are familiar with the first option. The first post in this series included an interactive section where you could test IP addresses. Stay tuned for IPv6. The numbers may seem random, but did you know your IP address matters? It matters for our privacy and could also matter to external actors if they can gain access to our systems. I, however, am talking about what it could reveal. Is your IP address private or public? How much does that affect the number of IP addresses that can be on a network? How are IP addresses organized into classes? In this post, I'll focus on answering each of these questions. Structure of an IPv4 Address Follows a dotted-decimal notation Has four octets Each octet has a value between 0 and 255, made up of 8 bits Has a total length of 32 bits in binary form What are I...
Post a Comment
Read more