MCP-demo

# MCP Integration This is a ** sampleproject** demonstrating how to connect an OpenAI Agent or other LLMs (with tool-calling capability) to an [MCP server](https://github.com/modelcontextprotocol) through langchain or openai-agent framework for enhanced paper search and chat functionality. It includes: - Creating and running an MCP server - Connecting an LLM (local or remote) to the MCP server - Running an interactive chatbot that can search arXiv papers and store results --- ## Usage Below are usage examples for each file script. --- ### `tool_function.py` It use arxiv lib to search and extract info from query of user. There are two function calling search and extract information. #### Run the test tool function ```sh poetry run python tool_function.py ``` ### `research_server.py` An example of creating a local MCP server with stdio connection. #### Run the test tool function ```sh npx @modelcontextprotocol/inspector run research_server.py ``` ## Serving Locally with vLLMs You can run this project with a local [vLLM](https://github.com/vllm-project/vllm) for faster inference. For GPT-OSS-20B, see [GPT-OSS instructions](https://docs.vllm.ai/projects/recipes/en/latest/OpenAI/GPT-OSS.html#installation). **Important:** The LLM you choose must have **tool-calling capability** to work with the MCP server. --- ### Start a Local vLLM Server Open new terminal then run. ```sh vllm serve openai/gpt-oss-20b ``` Then update `init_agent()` to point to your vLLM endpoint (default: `http://localhost:8000/v1`). --- ### `chatbot_mcp.py` ```sh python chatbot_mcp.py ``` Initializes an agent and connects to the MCP server. Each agent maintains its own session.