Skip to content

Prerequisites

Before configuring Claude Desktop, ensure that the following are installed:

  • Claude Desktop
  • Conda or Anaconda
  • The AIRA Conda environment
  • Node.js and npx for the Filesystem MCP server

Install Node.js and npx on macOS

The Filesystem MCP server is launched through npx. Check whether Node.js and npx are already available:

node --version
npx --version

If either command is unavailable, install Node.js with Homebrew:

brew install node

Verify the installation:

node --version
npm --version
npx --version

The AI Research Assistant MCP server runs through Python. Node.js and npx are required only for the Filesystem MCP server.

Connecting Claude Desktop with MCP

This guide explains how to connect Claude Desktop to the AI Research Assistant using the Model Context Protocol (MCP).

The project exposes a local MCP server that allows Claude Desktop to:

  • Create research databases from arXiv
  • List available databases
  • View database metadata
  • Ask research questions using the shared RAG pipeline

MCP Architecture

flowchart LR

    Claude["Claude Desktop"]

    subgraph MCP["MCP Servers"]
        FS["Filesystem Server"]
        AI["AI Research Assistant Server"]
    end

    subgraph Project["AI_RA"]
        DATA["data/"]
        VECTOR["vectorstore/"]
        TOOLS["RAG Tools"]
    end

    Claude --> FS
    Claude --> AI

    FS --> DATA
    FS --> VECTOR

    AI --> TOOLS
    TOOLS --> VECTOR

The two MCP servers provide complementary capabilities:

Server Purpose
Filesystem Allows Claude Desktop to access the project's data/ and vectorstore/ directories.
AI Research Assistant Exposes the RAG pipeline through MCP tools for creating databases and answering research questions.

Configure Claude Desktop

On macOS, open:

~/Library/Application Support/Claude/claude_desktop_config.json

Add the following configuration (replace the paths with your own):

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/Users/<username>/AI_RA/data",
        "/Users/<username>/AI_RA/vectorstore"
      ]
    },
    "ai-research-assistant": {
      "command": "/bin/zsh",
      "args": [
        "-lc",
        "cd /Users/<username>/AI_RA && /Users/<username>/anaconda3/envs/AIRA/bin/python -m mcp_server.server"
      ]
    }
  }
}

Save the file and restart Claude Desktop.

Available Tools

Once connected, Claude Desktop can invoke the following tools exposed by the AI Research Assistant:

Tool Description
list_databases List all available research databases.
build_database Create a new research database from arXiv papers.
database_status View metadata for a research database.
ask_research_question Query a selected research database using the local RAG pipeline.
delete_research_database Delete the selected research database and the respective downloaded papers.

Working with Research Databases

Research databases are shared between the Streamlit application and the MCP server.

For example:

  • A database created through Streamlit is immediately available in Claude Desktop.
  • A database created through Claude Desktop appears in the Streamlit interface.

Both interfaces operate on the same Chroma vector databases stored in the vectorstore/ directory.

Example Workflow

Ask Claude to create a new database:

Build a research database named Vision Transformers using the arXiv topic vision transformers and download 20 papers.

Then ask a question:

What positional encoding methods are commonly used in Vision Transformers?

Claude will automatically invoke the appropriate MCP tools, retrieve relevant document chunks from the selected database, and generate a grounded response using the local MLX language model.