Why MCP really is a big deal | Model Context Protocol with Tim Berglund

Understanding the Real Power of MCP

Model Context Protocol (MCP) has emerged as a foundational technology for modern AI systems, but its true significance often gets misunderstood. Most discussions focus on enhancing desktop applications with basic agentic features, treating MCP as a simple tool for adding AI-powered functionality to existing software. However, the real breakthrough lies in recognizing MCP as an architecture for exposing reusable tools and resources—including real-time data streams from sources like Apache Kafka—that enable the creation of sophisticated agentic microservices. This distinction is crucial: MCP transforms AI from a conversational layer into an operational system capable of autonomous problem-solving in professional environments.

The Limitations of Raw Language Models

Large Language Models are remarkable at generating text and reasoning through complex problems, but they operate within significant constraints. Without direct access to external tools and resources, LLMs cannot take meaningful actions in the world. They cannot schedule appointments, query databases, process transactions, or respond to real-time events. The model receives a prompt, generates a response, and the interaction ends—useful for information retrieval and creative tasks, but insufficient for systems that must act autonomously. Understanding these limitations is essential for architects designing AI systems that go beyond conversational interfaces to become true agents capable of executing workflows and managing business processes.

Tools and Resources as the Foundation of Agency

What distinguishes an agentic system from a simple chatbot is access to tools and resources that enable concrete action. An agentic AI system needs the ability to invoke functions, retrieve real-time information, modify data, and integrate with existing business systems. MCP provides the standardized mechanism for exposing these capabilities. Whether the tool is a calendar API for scheduling, a database connector for queries, or a streaming data source for real-time analytics, MCP creates a protocol-driven interface that allows AI models to discover, understand, and invoke these resources. This architectural approach ensures that tools can be composed, reused across multiple applications, and maintained independently from the AI logic consuming them.

The Three-Layer Architecture of MCP

MCP operates on a clean three-layer architecture that separates concerns and enables modular development. The host application is the entry point—a desktop tool, web service, or microservice that users interact with. The client library handles communication with MCP servers and manages the protocol implementation, translating high-level requests into standardized MCP calls. The server exposes the actual tools and resources: database connections, API endpoints, real-time data streams, business logic functions, and any other capability the system needs. This separation is powerful because it allows teams to develop tools independently, expose them through a standard protocol, and compose them into sophisticated workflows. A single MCP server can serve multiple host applications, and multiple servers can be chained to create complex agentic systems.

From Chatbots to Agentic Microservices

The conceptual leap from chatbot to agentic microservice represents a fundamental shift in how organizations deploy AI. A traditional chatbot responds to user queries with generated text. An agentic microservice receives a goal or problem statement, accesses relevant tools and data through MCP, and orchestrates a workflow to achieve that goal. Consider an appointment scheduling system: rather than a user typing requests and reading responses, the agentic system directly accesses calendar databases, understands availability constraints, communicates with other services, and completes the scheduling autonomously. This capability transforms AI from an interface layer into operational infrastructure capable of handling business processes end-to-end. MCP provides the protocol foundation that makes this possible by standardizing how agents discover and interact with the resources needed for real-world problem-solving.

Practical Example: Agentic Appointment Scheduling

A concrete example illustrates how MCP enables real-world agentic applications. An appointment scheduling system needs to access calendar data, understand meeting durations and constraints, check participant availability, handle conflicting requests, and confirm bookings. With MCP, the agentic system exposes these capabilities through dedicated tools: a calendar query tool that retrieves availability, a scheduling tool that commits bookings, and possibly an integration tool that synchronizes with external calendar services. When a user requests an appointment, the agent receives the request, uses MCP-exposed tools to query current availability, understands constraints through natural language processing, invokes the scheduling tool with appropriate parameters, and returns confirmation to the user—all without requiring manual intervention at each step. This demonstrates how MCP moves beyond theoretical agentic capabilities to deliver measurable value in professional settings.

The Complete Agentic Workflow Pattern

Effective agentic systems follow a consistent workflow enabled by MCP. The process begins with prompting: the agent receives an instruction or goal, often expressed in natural language. The LLM analyzes the prompt and determines which tools and resources are needed to progress toward the goal. The agent then accesses resources through MCP: querying databases, retrieving real-time data streams, checking current state. Based on the information retrieved, the agent invokes relevant tools to take action: updating records, triggering processes, coordinating with other systems. This cycle repeats until the goal is achieved or the agent determines that further action is impossible. MCP standardizes each step in this workflow, ensuring that different tools integrate seamlessly and that agents can reason about which resources to invoke based on their capabilities and the current task state.

Real-Time Data and Event-Driven Agentic Systems

One particularly powerful application of MCP is integration with real-time data sources like Apache Kafka. Traditional batch-based systems process data in windows—analyzing yesterday's transactions, generating reports weekly, updating forecasts monthly. Agentic systems integrated with real-time data sources can respond immediately to events. An agent monitoring a Kafka stream can detect anomalies as they occur, invoke tools to investigate and respond, and adjust behavior based on emerging patterns. MCP provides the mechanism for agents to subscribe to data streams, understand schema and semantics, and trigger actions based on real-time observations. This capability is transformative for domains like fraud detection, supply chain optimization, and dynamic resource allocation, where immediate response to events provides significant competitive advantage.