USB-C for Software MCP, the Universal Agent Interface

People in the software industry often claimed that things are moving fast, and things are constantly changing. I usually had the opinion that this is not true in general, because the core principles of software development and architecture were not changing too much.

Now, with the rise of AI and LLMs, I think we are actually at a point where we will see major changes. Most best preactices of software development and architecture still apply but the way that users will interact with software products will change and it is coming fast. I think within the next 12 months we will see a major shift and the reason is the introduction and adoption of MCP, the Model Context Protocol.

AI Agents are Becoming the Universal Interface

AI Agents like ChatGPT, Claude or Gemini are becoming the general-purpose interface layer for users. Already today, users increasingly prefer to interact with these agents when they look for information rather than using traditional search engines. A 2024 survey found that 42% of users report Google Search has become less useful, while Google CEO Sundar Pichai acknowledges that “People are using Search in entirely new ways, asking new types of questions — longer and more complex queries … and getting back the best the web has to offer.” as users turn to AI for more complex, conversational queries.

Especially, when we look at the capabilities of current models like OpenAIs o3 and o4-mini that can use tools like web search during their reasoning process and therefore produce more accurate and up-to-date results it gets clear how convenient this is for users.

This trend will continue and accelerate, leading to a future where users delegate tasks to AI agents instead of clicking through traditional search result pages, user interfaces and apps. Users will use natural language interfaces like chats or verbal converstations to interact with their AI agents.

Imagine users “talking” to their AI agent like this:

Notice what’s happening here: the user articulates their desired outcome, not a sequence of steps like they would do when they would use traditional apps. The AI agent orchestrates multiple services behind the scenes—traffic data, calendar systems, time tracking, CRM systems—while the user focuses on what they want to achieve.

This represents a fundamental shift from sequential app interactions to outcome-based delegation.

Where Agents Win — and Where GUIs Still Matter

However, it’s important to recognize that this shift won’t be uniform across all use cases. Jakob Nielsen, the “guru of web page usability”, wrote in a recent article:

“Once the transition to agents has been completed, websites may never see a human user again. This means that UI design, as traditionally conceived, becomes irrelevant. Look and feel? Bah, nobody will care whether it looks nice or ugly or how it feels to use your website since no (human) will be using it.”
— Jakob Nielsen

While Nielsen’s statement sounds radical, I believe it’s accurate for bounded, predictable, and repetitive tasks that can be fully automated by AI agents. But I think, traditional GUIs remain superior for creative, open-ended workflows—though these too will benefit from agent assistance and will evolve over time, they’ll never completely disappear.

Here are some examples to illustrate this point:

Pure agent replacement happens for routine, bounded tasks:

• “Will I be late for my morning meeting?” (replaces the need to open calendar, weather and map apps)
• “Schedule a recurring meeting when everyone’s available” (eliminates the need for complex calendar UIs)
• “Find me a tech conference next month within 3 hours from Berlin” (replaces browsing multiple event sites)

Agent-assisted creativity with simpler less complex interfaces:

• Agents that control Blender via MCP to generate 3D assets for game development
• Design agents that access your brand guidelines and generate initial layouts in Figma
• Video editing agents that analyze your footage and suggest cuts based on your past editing patterns
• Coding agents that generate code based on your coding style and project context

So the agent revolution isn’t a binary replacement of all user interfaces, but it is still a fundamental shift. Especially, and this is the key point, when the agents get full access to the products and services.

Currently, this isn’t widespread practice. Independent agents like ChatGPT primarily have access only to their training data and web search capabilities. Proprietary agents embedded in products like Notion, Microsoft Copilot, or Google Workspace can leverage their host product’s data and functionality, but they aren’t interoperable with each other or with third-party agents. Most companies running their own AI agents do so on proprietary infrastructure, which is both expensive and difficult to scale for smaller organizations.

The emerging solution is an open, interoperable protocol that enables agents to access products and services through standardized interfaces. This creates a clean separation between the agent’s reasoning capabilities and your product’s functionality, allowing agents to understand and leverage your offerings without needing to replicate your entire user interface.

MCP: The Missing Link for Agent Interoperability

This isn’t just a future vision—it’s happening today. Anthropic’s introduction of MCP (Model Context Protocol), combined with widespread industry adoption by major players like OpenAI and Microsoft, signals that we’ve reached a crucial tipping point. MCP serves as the critical bridge between AI agents and products and services, enabling seamless structured interaction.

Think of MCP as the USB-C port for AI agents—a standardized way for your product to communicate what it can do, what data it has, and how agents can interact with it effectively.

Traditional Apps and Web Services would look like this:

An MCP server can be implemented as a lightweight backend service that connects ot existing APIs like this:

AI agents connect to your product through MCP servers, either running locally on the user’s machine or accessible via web requests. Just as mobile apps use REST APIs today, agents use MCP servers to access your product’s capabilities. However, MCP goes beyond traditional APIs by adding structured metadata—tool descriptions, parameter schemas, and behavioral hints—that helps agents understand what your tools do and how to use them effectively. The integration is remarkably simple: users just add your MCP server endpoint to their agent’s configuration, and the agent immediately gains access to your product’s exposed capabilities.

An MCP server is a lightweight, easy-to-implement backend service that exposes your product’s capabilities through three core components:

• Resources File-like data that can be read by clients (like API responses or file contents)
• Tools Functions that can be called by the agent
• Prompts Pre-written templates that help users accomplish specific tasks

A great real-world example is the official GitHub MCP server, which allows AI agents to interact with repositories, issues, and pull requests through a standardized interface—turning GitHub’s existing API into an agent-accessible tool with minimal effort. Other examples include players like PayPal, Stripe, Atlassian and many more.

Current Limitations and Future Directions

MCP is very much in the early stages, and there are still some features and best practices missing. Like the smooth integration of authentication (already covered in the MCP standard and rolling adoptions happening) and authorization and patterns on how to implement the MCP server in a secure and policy compliant way.

Also, major Agent providers like OpenAI only recently added beta support for MCP, so it might look like there is a lot of uncertainty. While this is true, most missing features and upcoming changes are predictable and can be safely and sustainably worked around.

The direction is clear: We will see a significant shift from traditional user interfaces to AI agent-led interfaces, and structured protocols like MCP will be the standard way to expose products and services to AI agents. The specific protocols may evolve and change, but the basic principle remains: AI agents need structured, standardized ways to understand and interact with products and services.

So, put MCP on your radar and think about how you can prepare your product for this new paradigm. If you need help in getting started with MCP, feel free to reach out or subscribe to my newsletter or follow me on LinkedIn.