MCP Research Server (arXiv)
This project is a hands-on implementation of the Model Context Protocol (MCP) to demonstrate how an AI client can dynamically discover and call tools exposed by a server.
The server exposes research-related tools backed by arXiv, and the client uses an LLM (Mistral via Ollama) to decide when and how to call those tools.
This project is intentionally simple and educational, built to understand MCP fundamentals end-to-end.
What is MCP?
Model Context Protocol (MCP) is a standard that allows:
- A server to expose tools (functions + schemas)
- A client (LLM-powered) to discover available tools dynamically
- The LLM to decide when to call a tool
- Tool results to be fed back into the model for final reasoning
Think of MCP as a clean contract between:
- Tools
- Schemas
- LLM reasoning
- Execution
Project Architecture
MCP Research Server (arXiv)
This project is a hands-on implementation of the Model Context Protocol (MCP) to demonstrate how an AI client can dynamically discover and call tools exposed by a server.
The server exposes research-related tools backed by arXiv, and the client uses an LLM (Mistral via Ollama) to decide when and how to call those tools.
This project is intentionally simple and educational, built to understand MCP fundamentals end-to-end.
What is MCP?
Model Context Protocol (MCP) is a standard that allows:
- A server to expose tools (functions + schemas)
- A client (LLM-powered) to discover available tools dynamically
- The LLM to decide when to call a tool
- Tool results to be fed back into the model for final reasoning
Think of MCP as a clean contract between:
- Tools
- Schemas
- LLM reasoning
- Execution
Project Architecture
MCP Server (research_server.py)
The MCP server exposes two tools:
1. search_papers
Searches arXiv for papers by topic and stores metadata locally.
Inputs
- topic (string)
- max_results (int)
Returns
- List of arXiv paper IDs
2. extract_info
Fetches stored metadata for a given paper ID.
Inputs
- paper_id (string)
Returns
- Title, authors, abstract, PDF link, publish date
The server runs using stdio transport, which is ideal for local development and MCP inspectors.
MCP Client (mcp_chatbot.py)
The client:
- Connects to the MCP server over stdio
- Discovers available tools dynamically
- Uses Mistral (via Ollama) to decide:
- Whether a tool is needed
- Which tool to call
- What arguments to pass
- Calls the tool through MCP
- Feeds the tool result back to the LLM for a final response
The LLM never hardcodes tool logic — it reasons based on tool schemas.
LLM Used
- Model: Mistral
- Runtime: Ollama (local)
- Reason: Fast, local, great for experimentation without API costs
Make sure Ollama is running:
ollama serve
## Installation
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -r requirements.txt
## Running MCP server
python research_server.py
## Running MCP client
python mcp_chatbot.py
## Example interaction
Query: Can you search papers around physics and find 2 of them?
Searching arXiv for 'physics' (2 papers)...
Papers found:
- 1910.11775v2
- hep-ex/9605011v1