The Model Context Protocol, commonly referred to as MCP, is an open protocol developed by Anthropic that enables AI agents and assistants to access tools and resources in a standardized way.
Before you learn more about MCP, it is important to understand the concept of an Agent.
What are Agents?
There are many definitions of an Agent, but for the context of this course, an Agent is a system that acts independently, using tool calling to access information and take actions to achieve a specific goal.
A common type of framework for building agents is the ReAct framework (Reason/Act). When a ReAct agent receives an input, it follows a continuous loop of planning, reasoning, and acting until the goal is achieved.
The process begins with planning, where the LLM analyzes the task, breaks it down into sub-tasks, extracts required information and parameters, and may elicit additional details from the user if needed. During this phase, the agent determines what tools will be needed and in what sequence they should be called.
Next comes reasoning, where the LLM uses the descriptions of available tools to select the correct tool(s) for each sub-task. The agent evaluates which tools are most appropriate based on the current context and requirements.
Finally, the agent acts by calling the selected tools, either in sequence or in parallel, to gather information and perform the necessary actions. The results from these tool calls are then fed back into the planning phase, creating a continuous loop where the agent refines its approach based on new information.
For example, consider a smart home automation agent, that uses the weather forecast combined with local temperature sensor data to adjust the blinds and turn on an air conditioning unit.
Where does MCP fit in?
While agents may have access to tools written directly into their codebase, MCP provides a standardized way for the agent to consume tools built by third parties. An agent could be configured to access a weather API service hosted in the cloud, or connect to a local database tool running on the same machine through Standard Input/Output (stdio).
Architecture
The MCP protocol consists of the following elements:
-
Servers
-
Clients
-
Hosts
Servers
The protocol follows a client-server architecture, similar to microservices, where Servers provide capabilities through a list of tools. Each tool that the server hosts has additional metadata assigned to it, including a unique identifier, a description of when that tool can be used, and a list of parameters that the tool can receive. Servers describe the parameters that the tool can receive, and the expected output of the tool.
Along with tools, servers provide access to Resources, read-only data that can be consumed by the client. Each resource has its own unique URI, similar to an endpoint in a REST API, and can be used to provide direct access to specific objects with a fixed location.
MCP tools can also provide Prompt templates, allowing the client to use pre-written prompts that follow best practices.
Clients
A Client is responsible for managing a one-to-one connection to a server.
When a client starts, it will connect to a server, and request a list of tools, resources, and prompt templates that the server provides access to.
This is a stateful connection, meaning the client maintains an ongoing connection to the server throughout its lifecycle. The connection can be established using different transport methods:
-
stdio - The client starts the server locally and communicates through standard input/output streams
-
HTTP - The client connects to a remotely hosted server using HTTP Server Sent Events (SSE) or HTTP streaming
You will learn more about transport methods in the next lesson.
Hosts
One or more clients are managed by a Host, an application or piece of software that maintains session state and context.
An example of a host may be Claude Desktop, Amazon Q, VS Code, or Cursor. Your agent applications are also examples of an MCP host.
The host determines which tool to use and provides the appropriate parameters to the tool. The client then executes the tool on behalf of the host.
Tools
Each tool will have a unique identifier, a description of when that tool can be used, and a list of parameters that the tool can receive.
For example, a tool for searching the web may be defined as:
-
id: search-web
-
description: Search the web for information
-
parameters: query, a string
Tools can output a result as text or return Base-64 encoded images.
Lesson summary
In this lesson, you learned:
-
Agents are systems that act independently to achieve specific goals, often using the ReAct framework (reasoning and acting with tools)
-
MCP (Model Context Protocol) provides a standardized way for agents to consume tools from external providers
-
MCP Architecture consists of four key components:
-
Servers provide capabilities through tools, resources, and prompt templates
-
Clients manage one-to-one connections to servers
-
Hosts (like Claude Desktop) manage clients and decide which tools to use
-
Tools have unique identifiers, descriptions, and parameters
-
-
Real-world usage involves configuring hosts to connect to MCP servers for enhanced functionality
You can find more information about these terms at modelcontextprotocol.io.