Overview

Nexarag is an open-source platform for building knowledge graphs from research papers and querying them with AI, enabling transparent and reproducible literature analysis without the hallucinations of traditional RAG systems. Deploy locally with full privacy control or integrate with any LLM via the standardized Model Context Protocol (MCP).

See also:

• Usage and Development Guides

Feedback

We are actively seeking feedback for Nexarag, including feature requests, issue reports, training material, etc. Please submit to nexarag.ai@gmail.com.

Please submit any bugs to the issues page.

Test Coverage

• Test Coverage Report

Pre-Requisites

• Docker
• (Windows Only) WSL2
• (MacOS Only) Ollama Desktop
• (Optional) Claude Desktop (for MCP)

Quick Start

1. Download Docker compose file

Choose a Dockerfile compatible with your OS and hardware:

• MacOS
• Linux/WSL (CPU)
• Linux/WSL (GPU)

[Optional]: Move the Docker compose file to a location on your drive, e.g. ~/Nexarag.

2. Run Nexarag

From the same directory as the downloaded docker-compose.yml, run:

docker compose up -d

Visit Nexarag in your browser at http://localhost:5000.

3. Pull Ollama Models

To support all internal features, Nexarag requires:

• An embedding model, such as nomic-embed-text:v1.5
• A language model, such as gemma3:1b
• An MCP-capable model, such as qwen3:8b

Browse the full library of Ollama models here and choose any model from these families that your hardware supports. Defaults that run on most hardware are provided below.

Linux/WSL

Models can be pulled through the command line in the ollama Docker container.

docker exec -it nexarag.ollama /bin/bash
ollama pull nomic-embed-text:v1.5
ollama pull gemma3:1b
ollama pull qwen3:8b

MacOS

Pull models directly from your command line.

ollama pull nomic-embed-text:v1.5
ollama pull gemma3:1b
ollama pull qwen3:8b

4. (Optional) MCP Integration

Claude Desktop

Add the following to your Claude Desktop Config:

{
    "mcpServers": {
        "nexarag": {
            "command": "npx",
            "args": [
            "-y",
            "mcp-remote",
            "http://localhost:9000/mcp"
            ],
            "env": {
            "MCP_TRANSPORT_STRATEGY": "http-only"
            }
        }
    }
}

Ollama

First install pipx, then run:

pipx install ollmcp

To start the MCP client:

ollmcp -u http://localhost:9000/mcp -m qwen3:8b

Semantic Scholar

Please note that we are rate-limited by the Semantic Scholar API, so enriching BibTex uploads with data and updating the graph after adding papers from a Semantic Scholar search may take several minutes to complete.