llm-mcp-server-template

# llm-mcp-server-template Template project for LLM-MCP server development - The server provides the capability to compare numbers. - The client includes various models and calling methods. ## MCP - https://modelcontextprotocol.io/introduction - In simple terms, - The model (client) is the brain, responsible for planning and executing text-based tasks, such as writing code in the first step, executing it in the second step, and analyzing it in the third step. - MCP (Server) is the tool that provides additional capabilities needed by the model, such as mathematical calculations, file reading and writing, network requests, etc. ### MCP-Server - Primarily responsible for implementing MCP capabilities, such as mathematical calculations, file reading and writing, network requests, etc. - https://modelcontextprotocol.io/quickstart/server - During development, you can use the mcp-server-inspector tool to assist with debugging. - Types of servers: - 1. Local Stdio, where the client directly calls the local server file. - 2. Remote aggregation platforms using SSE, such as Glama.ai, MCP.so, etc. ### MCP-Client - Mainly responsible for coordinating the model and the MCP server, ultimately achieving "the model calls MCP capabilities." - Types of clients: - 1. Directly use existing client software, such as Cursor, MCP Inspector, etc. - 2. Various intelligent agents on cloud platforms, such as Alibaba Cloud's Bai Lian application management. - 3. Write your own scripts to support: - 3.1. The model's SDK already supports MCP, making it easy to call using scripts. - For example, Claude, https://modelcontextprotocol.io/quickstart/client - For example, OpenAI, https://openai.github.io/openai-agents-python/mcp/ - 3.2. The model's SDK does not support MCP, requiring custom implementation. - 3.2.1. Use frameworks, such as langchain-mcp-adapter (https://github.com/langchain-ai/langchain-mcp-adapters) - 3.2.2. Fully native, specifically including multi-round requests: - a. Use a prompt to request the model, informing it of the task and introducing the MCP servers, and the model returns the interface and parameters to call the server. - b. Use the information returned from the previous round to request tools and obtain results. - c. Use the information returned from the previous round to request the model and obtain results. - d. Repeat b and c until the model returns the final result. ## Project Related ### Environment Preparation ``` # Install the uv package management tool, skip if already installed curl -LsSf https://astral.sh/uv/install.sh | sh source ~/.local/bin/env uv --version # Install Python 3.10, skip if already installed uv python list uv python install 3.10 cd server # In the server directory, create a virtual environment, skip if already created uv venv --python 3.10 source .venv/bin/activate uv add "mcp[cli]" httpx # In the server directory, use mcp-server-inspect for debugging during development mcp dev math.py cd client # In the client directory, create a virtual environment, skip if already created uv venv --python 3.10 source .venv/bin/activate uv add openai-agents socksio touch .env # In the .env file, set OPENAI_API_KEY, etc. # In the client directory, run normally source .venv/bin/activate python openai_client.py ``` ### Project Results MCP Server, mcp-server-inspect debugging page:  MCP Client, ... ## TODO ~~1. mcp-server, local development and debugging~~ 2. mcp-server, publish remote hosting platform 3. mcp-client, use openai-agent-sdk to call the server 4. mcp-client, use anthropic-sdk to call the local server 5. mcp-client, use langchain-mcp-adapter to call the server