NerdCabalMCP

ARTIFEX NERD SWARM - Quick Start Guide

🎯 Overview

The ARTIFEX NERD SWARM is a Model Context Protocol (MCP) server providing 11 specialized AI agents for research operations. This guide shows you how to install, run, and use the system.

📋 Prerequisites

Optional (for specific agents):

🔧 Installation

Step 1: Clone the Repository

git clone https://github.com/Tuesdaythe13th/Paper2Agent.git
cd Paper2Agent/mcp-server

Step 2: Install Dependencies

npm install

This installs the MCP SDK and TypeScript dependencies.

Step 3: Build the Project

npm run build

This compiles TypeScript to JavaScript in the dist/ directory.

🚀 Running the MCP Server

1. Locate Claude Desktop Config

macOS:

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

Windows:

%APPDATA%\Claude\claude_desktop_config.json

Linux:

~/.config/Claude/claude_desktop_config.json

2. Edit the Config File

Add the ARTIFEX NERD SWARM server to your MCP servers:

{
  "mcpServers": {
    "artifex-nerd-swarm": {
      "command": "node",
      "args": ["/absolute/path/to/Paper2Agent/mcp-server/dist/index.js"],
      "env": {}
    }
  }
}

Important: Replace /absolute/path/to/Paper2Agent with your actual path!

Example (macOS/Linux):

{
  "mcpServers": {
    "artifex-nerd-swarm": {
      "command": "node",
      "args": ["/home/user/Paper2Agent/mcp-server/dist/index.js"],
      "env": {}
    }
  }
}

Example (Windows):

{
  "mcpServers": {
    "artifex-nerd-swarm": {
      "command": "node",
      "args": ["C:\\Users\\YourName\\Paper2Agent\\mcp-server\\dist\\index.js"],
      "env": {}
    }
  }
}

3. Restart Claude Desktop

Quit and reopen Claude Desktop. The MCP server will start automatically.

4. Verify Connection

In Claude Desktop, you should see a 🔌 icon or MCP indicator showing “artifex-nerd-swarm” is connected.

Option 2: Run Standalone (for Testing)

You can run the server directly from the command line:

cd /path/to/Paper2Agent/mcp-server
node dist/index.js

The server runs on stdio (standard input/output), waiting for JSON-RPC 2.0 messages.

To test manually, you can send JSON-RPC requests:

echo '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | node dist/index.js

🎨 Using the Agents

Available Tools (14 total)

Once connected to Claude Desktop, you can use these MCP tools:

Tool Name Agent Purpose
create_paper_agent Infrastructure Create Paper2Agent from GitHub repo
get_pipeline_status Infrastructure Check pipeline progress
generate_rubric LLM-Rubric Architect Generate evaluation rubrics
design_experiment Experimental Designer Design research methodology
analyze_budget Budget Agent Budget analysis (grant/investor)
generate_comptroller_notebook The Comptroller Financial operations dashboard
generate_administrator_notebook The Administrator Organizational governance
generate_mlflow_query_plan MLflow MCP Agent Experiment tracking queries
generate_dataset_builder_plan Dataset Builder Training data construction
generate_ciso_notebook The CISO Security operations (STRIDE, incident response)
analyze_swarm_and_suggest The Orchestrator Swarm optimization & ADK workflows
create_design_system The Creative Director Brand/design system generation
visualize_dataset The Visual Inspector FiftyOne dataset visualization
analyze_transcript The Forensic Analyst Neural forensics with DSMMD taxonomy

📝 Example Usage

Example 1: Generate an Evaluation Rubric

In Claude Desktop:

Use the generate_rubric tool to create an evaluation rubric for:
- Task: Binary classification of medical images
- Domain: Radiology (chest X-rays)
- Metrics: Sensitivity, specificity, AUC-ROC
- Criteria: Clinical accuracy, fairness across demographics

Output: Google Colab notebook with evaluation framework

Example 2: Design an Experiment

In Claude Desktop:

Use the design_experiment tool to design a study on:
- Research question: Does data augmentation improve model robustness?
- Hypothesis: Augmented models will have 10% higher accuracy on corrupted images
- Dataset: CIFAR-10
- Methods: Train with/without augmentation, test on CIFAR-10-C

Output: Detailed experimental design with controls, variables, validation strategy

Example 3: Visualize Dataset Quality

In Claude Desktop:

Use the visualize_dataset tool with:
- mode: quality_check
- dataset: FiftyOne Zoo "quickstart" dataset
- brain method: mistakenness
- Check for label mistakes using model predictions

Output:

Example 4: Create a Design System

In Claude Desktop:

Use the create_design_system tool with:
- Project: ARTIFEX NERD SWARM Dashboard
- Aesthetic: functional_brutalist
- Modules: style_engine, interface_architect
- Components: dashboard, card
- Accessibility: WCAG_AA

Output:

Example 5: Security Review (CISO)

In Claude Desktop:

Use the generate_ciso_notebook tool to:
- Model threat scenarios: nation_state, ransomware_gang
- High-value assets: Patient health records, ML model weights
- Attack surfaces: Web API, Database, File storage
- Operational mode: offensive (red team simulation)

Output:

Example 6: Forensic Transcript Analysis

In Claude Desktop:

Use the analyze_transcript tool with:
- transcript source: inline
- content: [Your LLM conversation transcript]
- specimen_name: "Sediment/Juno"
- model_family: "GPT-4o"
- Enable all DSMMD detectors (110.1, 140.1, 140.3, 155.2, SB-1)

Output:

🧪 Testing Individual Agents

Quick Test Script

Create a test file test-agent.mjs:

import { generateRubricArtifacts } from './dist/rubric-architect.js';

const request = {
  taskDescription: "Evaluate sentiment analysis models",
  evaluationDomain: "Natural Language Processing",
  performanceMetrics: ["Accuracy", "F1-score", "Precision", "Recall"],
  qualityCriteria: ["Robustness to typos", "Multilingual support"]
};

const output = generateRubricArtifacts(request);
console.log(JSON.stringify(output, null, 2));

Run it:

node test-agent.mjs

🐛 Troubleshooting

Problem: “MCP server not connecting”

Solution:

  1. Check the config path is absolute (not relative)
  2. Verify dist/index.js exists: ls dist/index.js
  3. Check Node.js version: node --version (should be ≥18)
  4. Restart Claude Desktop completely

Problem: “Module not found” errors

Solution:

cd mcp-server
npm install
npm run build

Problem: “TypeError: Cannot read property…”

Solution: You may have old build artifacts. Clean and rebuild:

rm -rf dist/
npm run build

Problem: “JSON-RPC parse error”

Solution: The MCP client is sending malformed requests. Check your Claude Desktop config JSON syntax.

📚 Advanced Usage

Running Multiple Agents in a Workflow

Example: Budget-Constrained Experiment Design

1. Use design_experiment to create initial methodology
2. Use analyze_budget to estimate costs
3. If over budget:
   - Use design_experiment again with cost constraints
   - Repeat until budget satisfied
4. Use generate_ciso_notebook to check security implications
5. Use create_design_system to format final report

This demonstrates ADK LoopAgent pattern (iterative refinement).

Generating Notebooks for Google Colab

Most agents return notebooks in their output. To use them:

  1. Copy the notebook JSON from the agent output
  2. Create a new file notebook.ipynb
  3. Paste the JSON into the file
  4. Upload to Google Colab:
    • Go to https://colab.research.google.com
    • File → Upload notebook
    • Select your notebook.ipynb
  5. Run the cells interactively

Using with Other MCP Clients

The server follows the MCP specification, so it works with any MCP client:

Example: MCP Inspector (Testing Tool)

npm install -g @modelcontextprotocol/inspector

mcp-inspector node /path/to/Paper2Agent/mcp-server/dist/index.js

This opens a web UI to test tools interactively.

📖 Documentation

🔄 Development Workflow

Making Changes to Agents

  1. Edit TypeScript source: 
    cd mcp-server/src
vim creative-director.ts  # or your editor
  2. Rebuild: 
    npm run build
  3. Restart Claude Desktop to reload the MCP server

Adding a New Agent

  1. Create type definitions: src/my-agent-types.ts
  2. Create implementation: src/my-agent.ts
  3. Export AgentCard: export const MY_AGENT_CARD: AgentCard = {...}
  4. Add to index.ts:
    • Import: import { generateMyAgentOutput, MY_AGENT_CARD } from './my-agent.js';
    • Add tool definition to tools array
    • Add handler in switch (name) statement
  5. Rebuild: npm run build

🎯 Quick Start Checklist

📞 Support

Issues: https://github.com/Tuesdaythe13th/Paper2Agent/issues

MCP Documentation: https://modelcontextprotocol.io

FiftyOne (Visual Inspector): https://voxel51.com/docs/fiftyone

Happy researching with ARTIFEX NERD SWARM! 🚀

This site is open source. Improve this page.