Understanding MCP and Function Call: A Comprehensive Guide to LLM Tool Integration

The article begins by introducing the widely circulated MCP diagram and explains that discussing MCP inevitably involves Function Call, which is the fundamental protocol for large‑model tool usage and serves as a more complete engineering solution.

Function Call allows a model to select the appropriate tool from a provided list, extract correct parameters, and pass the request to a container for execution; the tool's result is then fed back as the next round of model context.

OpenAI's standard Function Call API is shown with a full JSON request example, illustrating the model , messages , and tools fields, as well as the required function schema for a query tool.

A concrete weather‑query example demonstrates how a user request is transformed into a prompt that includes tool signatures inside <tools> tags and how the model returns a JSON {"name":"getWeather","arguments":{...}} wrapped in <tool_call> tags.

The concept of special tokens such as <|im_start|> and <|im_end|> is introduced, explaining that they are learned during post‑training and act like markers in a story for a child.

Comparisons between traditional Function Call and newer approaches like CodeAct and MCP are discussed, highlighting that CodeAct pushes more work to the model (generating code) while MCP adds a client‑server layer to standardize tool invocation across different LLM providers.

The MCP architecture is broken down into three elements—host, client, and server—illustrated with examples from Ant Group, Anthropic, and Cursor, and the article notes that the host initiates a client which communicates with a server that abstracts tool differences.

Practical fast‑MCP Python code is provided, showing how to install the library, define tools with @mcp.tool() and resources with @mcp.resource() , and implement simple weather‑lookup functions.

For Java developers, the article presents a Spring MVC integration using the sofa-ai-mcp-server-webmvc-sofa-boot-starter dependency, a @Configuration class that registers WeatherTools as beans, and a client configuration that sets OpenAI endpoints, model options, and MCP client settings.

Finally, a complete Spring Boot application example demonstrates how to invoke the MCP‑enabled server from a command‑line runner, sending a weather query and printing the assistant's response.