mcp-client-demo

# MCP Client ## Content - Implement a basic Client to connect DeepSeek to the MCP ecosystem - Core Architecture - JSON-RPC 2.0 - What is RPC - **Differences between MCP and Function Calling** - API differences among various providers - OpenAI / Anthropic / DeepSeek - **Refer to 5ire's source code on how to unify API differences** ## MCP Architecture https://modelcontextprotocol.io/docs/concepts/architecture https://www.jsonrpc.org/ ## MCP Client Implementation https://modelcontextprotocol.io/quickstart/client https://github.com/modelcontextprotocol/typescript-sdk - MCP Server: https://github.com/pskill9/hn-server - Inspector: https://github.com/modelcontextprotocol/inspector ## Differences between MCP and Function Calling ### MCP Tool Invocation ```mermaid graph LR LLM <--> MCP_Client <--> MCP_Server ``` - MCP Client and MCP Server have a strict communication protocol - Server provides unified interface specifications and service registration mechanisms - Client knows how to communicate with Server and available service lists - Due to **unified protocol specifications**, tools developed by different developers can be easily reused by other MCP clients - Server's registration mechanism makes tool discovery and sharing simple ### Function Calling Tool Invocation ```mermaid graph LR LLM <--> AnyClient <--> AnyServer ``` - More like a **basic capability** of the model, LLM only identifies the function to be called - Client decides how to implement this function call - No unified service discovery and communication standards - Different clients' plugin ecosystems are not compatible with each other - https://chat-plugin-sdk.lobehub.com/quick-start/get-start - https://www.librechat.ai/docs/development/tools_and_plugins ## API Differences https://platform.openai.com/docs/api-reference/chat https://docs.anthropic.com/en/api/messages https://api-docs.deepseek.com/api/create-chat-completion - Differences 1. Function definition format for function calling 2. Data structure for function calling return 3. Message data structure after calling ## 5ire's Approach to Unifying API Differences - APIs of different models return different data structures, how to unify the output of LLM models and MCP-client differences? - Since we know they are different → we can convert them to be the same - https://github.com/nanbingxyz/5ire 1. Different function definition methods → ChatService's makeTool handles 2. Different function call results → Reader's parseReply and parseTools handle 3. Different message formats → ChatService's makeToolMessages handles