Understanding Model Context Protocol (MCP): Architecture, Execution Flow, and Ecosystem

1. MCP Technology

1.1 What is MCP

MCP (Model Context Protocol) is a standardized protocol that defines how applications exchange context information with AI models, acting as an intermediate layer similar to USB‑C, aiming to create a universal standard that simplifies and unifies AI application development and integration.

1.2 Core Components

The MCP architecture consists of MCP Server, MCP Client, and MCP Host, which communicate via defined interaction protocols.

1.2.1 MCP Server

The MCP Server is a lightweight application that provides tools execution, resource access, and predefined prompts to clients.

Resources are file‑like data, Tools are functions callable by the LLM, and Prompts are templates that guide specific tasks.

An example of a filesystem‑operation MCP Server is provided ( link ).

1.2.2 MCP Client

The MCP Client resides inside the Host application, maintaining a 1:1 connection with the MCP Server and acting as a bridge between the LLM and the server.

1.2.3 MCP Host

The Host is the LLM‑driven application (e.g., Claude Desktop, IDEs) that initiates requests through the MCP Client.

1.3 Communication Protocols

MCP uses JSON‑RPC encoded messages exchanged over standard input/output streams or Streamable HTTP.

1.3.1 Standard Input/Output (stdio)

Suitable for communication on the same machine.

1.3.2 Server‑Sent Events (SSE)

Enables real‑time data transfer across networks, useful for remote resources or distributed deployments.

1.4 Changes Brought by MCP

Before MCP, each intelligent application had to implement its own integration with external services. MCP standardizes the interface, allowing service providers to expose resources as MCP Servers, which can be consumed uniformly by applications, improving reuse and standardization.

2. MCP Execution Details

When a user submits a query in a Host application, the following steps occur:

MCP Client retrieves the list of available tools from the MCP Server.

The query and tool descriptions are sent to the LLM via function calling.

The LLM decides whether and which tools to use.

If tools are needed, the MCP Client invokes them through the MCP Server.

Tool results are returned to the LLM.

The LLM generates a natural‑language response.

The response is presented to the user.

2.1 Host Calls LLM

The Host sends the user query together with the tool list to the LLM.

2.2 Client Calls MCP Server

The LLM, based on the query and tool descriptions, may invoke the MCP Server for tool execution.

2.3 Final Result

The Host combines the MCP Server’s result with the LLM’s output and returns the final response to the user.

3. Agents, Function Calling and MCP

3.1 Agents

LLM agents are advanced AI systems that can reason, recall conversation history, and dynamically plan and execute tasks by invoking external tools.

3.2 Function Calling

OpenAI introduced function calling to give LLMs the ability to call external tools, enabling complex task automation through carefully designed prompts.

3.3 MCP

MCP addresses the inconsistencies of function calling by standardizing the interface, protocol, and implementation for agents to interact with external resources and tools.

3.4 Comparison

(Comparison diagram omitted.)

4. MCP Ecosystem

The release of Manus accelerated MCP adoption, leading to a growing marketplace of MCP Servers and applications.

4.1 MCP Server Marketplace

Various marketplaces (Smithery, mcp.so, PulseMCP, Cursor Directory, Glama, modelcontextprotocol, Cline) have seen rapid growth in the number of published MCP Servers within a short period.

4.2 MCP Applications

Over 200 applications on PulseMCP support MCP, covering many use cases.

4.3 MCP Server Types

MCP Servers now span data, research, cloud platforms, databases, chatbots, file systems, automation, and more. SDKs and frameworks such as FastMCP, Spring AI, and mcp‑agent lower development costs.

5. Reflections on MCP Development

5.1 Potential Changes

Analogous to how WebService, REST, and micro‑services reshaped system integration, MCP may become an OpenAPI‑like protocol for AI services, reshape application interaction, enable end‑to‑end business automation, and foster open ecosystem collaboration.

5.2 Technical Limitations

Current challenges include limited application scope (many servers run locally), competition with other standards, reliance on Anthropic’s Claude ecosystem, questions about whether MCP is a true protocol, and security risks from locally deployed servers gaining OS access.