Ollama + MCP: Local AI Research Assistant

Ever wanted to chat with your local AI about research papers without sending data to external APIs? This project lets you do exactly that using Ollama and the Model Context Protocol (MCP).

What it does

I built this to learn more about how MCP works and how to get it working with Ollama. Turns out it's also pretty useful for research - now I can just ask my local AI model to find papers, summarize them, and keep everything organized automatically.

The setup includes:

• Custom ArXiv research server - searches and organizes academic papers (built following an MCP course)
• File system access - using the official MCP filesystem server
• Web scraping - using the official MCP fetch server
• MCP client that works with Ollama - the tricky integration part I figured out

Quick start

You'll need Python 3.13+, UV package manager, and Ollama running locally.

# Get the code
git clone https://github.com/career-genomics/ollama-mcp.git
cd ollama-mcp

# Set up environment  
uv venv && source .venv/bin/activate
uv pip install -e .

# Make sure Ollama is running
ollama serve
ollama pull qwen3:14b  # or qwen3:8b for lighter usage

How to use it

Full MCP experience (recommended)

python src/mcp_client/mcp_chatbot.py

Then you can:

• Ask: "Find papers about quantum machine learning"
• Use @folders to see your research topics
• Use @quantum_machine_learning to browse saved papers
• Regular research questions work great too

Simple chatbot (no MCP setup needed)

python src/chat-bot/chatbot-loop.py

Just the ArXiv server

python main.py

What makes it useful

The AI can actually do things, not just chat:

• Search ArXiv and save papers organized by topic
• Remember what it found in previous sessions
• Fetch web content and analyze it
• Work with your local files

All your data stays local. The AI runs on your machine through Ollama, and research papers get saved in a simple papers/ folder structure.

Configuration

Want to use a different model? Edit the model name in src/mcp_client/mcp_chatbot.py:

chatbot = MCPChatBot(model_name="qwen3:8b")  # or llama2, etc.

The MCP servers are configured in src/configs/server_config.json - you can add or remove services there.

Example workflow

You: "Find recent papers on transformer improvements"

AI: [searches ArXiv, finds 5 papers, saves them locally]
    "I found 5 recent papers on transformer improvements. Here's what I discovered..."
    [provides detailed analysis]

You: "@folders"
AI: Shows: transformer_improvements

You: "@transformer_improvements"  
AI: [displays organized information about all saved papers]

Project structure

src/
  mcp_client/mcp_chatbot.py     # Full-featured MCP client
  mcp_servers/arxiv_research/   # ArXiv search server
  chat-bot/chatbot-loop.py      # Simple standalone version
  configs/server_config.json    # MCP server settings
papers/                         # Where research gets saved

Common issues

Can't connect to Ollama? Make sure it's running with ollama serve

MCP servers not working? Check that all dependencies installed with uv pip install -e .

Out of memory? Try a smaller model like qwen3:8b instead of the 14B version

Why I built this

I couldn't find a consistent, working way to run MCP servers with Ollama. Most examples online are either incomplete, outdated, or assume you're using Claude/OpenAI APIs.

The official MCP servers (filesystem, fetch, etc.) work great, but connecting them to Ollama models was the missing piece. After digging through docs and experimenting, I figured out how to make the integration work reliably - handling tool calling format conversion, message flow, and multi-server coordination.

I also built a custom ArXiv research server since I was tired of manually tracking papers. Now it's all automated and stays local.

Contributing

Found a bug or want to add a feature? Pull requests welcome! The code is pretty straightforward to understand and modify.

The filesystem and fetch servers come from the official MCP servers repository - the main contribution here is the Ollama integration layer and custom ArXiv research server.

License

Apache 2.0 - use it however you want.

---

Built because research should be easier, not harder.