AsterMind-ELM MCP Server

Fast, on-device machine learning with production-grade persistence through the Model Context Protocol.

Overview

This MCP server brings the power of AsterMind-ELM to Claude with full persistence capabilities, enabling:

• Instant Training - Train text classifiers in milliseconds using Extreme Learning Machines
• Production Persistence - Models and datasets survive server restarts via SurrealDB
• Performance Monitoring - Track accuracy, drift, and prediction logs over time
• On-Device Processing - All computation happens locally, no external APIs
• Privacy-First - Your data never leaves your machine
• Microsecond Inference - Real-time predictions
• Vector Storage - Embedding storage with similarity search

Key Features

• 🚀 Fast Training: Extreme Learning Machines use closed-form solutions (no gradient descent)
• 💾 Production-Ready Persistence: SurrealDB backend for models, datasets, and monitoring
• 📊 Model Monitoring: Track performance metrics, confusion matrices, and drift detection
• 🔒 Private: All processing happens on-device
• 🎯 Versatile: Classification, embeddings, online learning, similarity search
• 📈 Reproducible: Link models to datasets for full experiment tracking

Installation

1. Install Dependencies

cd /path/to/astermind-elm-mcp
npm install

2. Set Up SurrealDB (Optional but Recommended)

Install SurrealDB:

# Linux
curl -sSf https://install.surrealdb.com | sh

# macOS
brew install surrealdb/tap/surreal

# Or download from https://surrealdb.com/install

Start SurrealDB:

surreal start --log trace --user root --pass root memory

Initialize the database schema:

npm run build
node build/scripts/init-db.js

3. Build the Server

npm run build

4. Configure Claude Desktop

Edit your Claude Desktop config file:

• Linux: ~/.config/Claude/claude_desktop_config.json
• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

Add the server configuration:

{
  "mcpServers": {
    "astermind-elm": {
      "command": "node",
      "args": [
        "/full/path/to/astermind-elm-mcp/build/index.js"
      ],
      "env": {
        "ENABLE_PERSISTENCE": "true",
        "LOG_PREDICTIONS": "true",
        "SURREALDB_URL": "ws://127.0.0.1:8000/rpc",
        "SURREALDB_NAMESPACE": "astermind",
        "SURREALDB_DATABASE": "production",
        "SURREALDB_USERNAME": "root",
        "SURREALDB_PASSWORD": "root"
      }
    }
  }
}

5. Restart Claude Desktop

Environment Variables

Variable	Default	Description
ENABLE_PERSISTENCE	false	Enable SurrealDB persistence
LOG_PREDICTIONS	false	Auto-log all predictions for monitoring
SURREALDB_URL	ws://127.0.0.1:8000/rpc	SurrealDB connection URL
SURREALDB_NAMESPACE	astermind	Database namespace
SURREALDB_DATABASE	production	Database name
SURREALDB_USERNAME	root	Database username
SURREALDB_PASSWORD	root	Database password

Available Tools

Core Training & Inference

train_classifier

Train a text classification model using Extreme Learning Machine.

Parameters:

• model_id (string, required): Unique identifier for the model
• training_data (array, required): Array of {text: string, label: string} objects
• config (object, optional): Classifier configuration
  • hiddenUnits (number): Hidden layer size (default: 128)
  • activation (string): relu, leakyrelu, sigmoid, tanh, linear, gelu (default: relu)
  • weightInit (string): uniform, xavier, he (default: xavier)
  • ridgeLambda (number): Ridge regularization (default: 1e-6)
  • maxLen (number): Max sequence length (default: 30)
  • dropout (number): Dropout rate (default: 0)
• description (string, optional): Model description
• persist (boolean, optional): Save to database immediately
• version (string, optional): Version string (default: timestamp)
• dataset_id (string, optional): Link to stored dataset
• tags (array, optional): Tags for organization

Example:

{
  "model_id": "sentiment_v1",
  "training_data": [
    {"text": "I love this!", "label": "positive"},
    {"text": "This is terrible", "label": "negative"}
  ],
  "persist": true,
  "version": "1.0.0",
  "tags": ["sentiment", "production"]
}

predict

Make predictions using a trained model.

Parameters:

• model_id (string, required): Model identifier
• text (string, required): Text to classify
• top_k (number, optional): Number of predictions (default: 3)
• log_prediction (boolean, optional): Log to database for monitoring
• ground_truth (string, optional): True label for accuracy tracking

generate_embedding

Generate embedding vector from text.

Parameters:

• model_id (string, required): Model to use
• text (string, required): Text to embed

list_models

List all models currently in memory.

delete_model

Remove a model from memory.

Parameters:

• model_id (string, required): Model to delete

save_model

Export model summary (in-memory models only).

Parameters:

• model_id (string, required): Model to export

Persistence Tools

store_model_persistent

Save a trained model to SurrealDB. Model survives server restarts.

Parameters:

• model_id (string, required): Model in memory to persist
• version (string, required): Version string (e.g., "1.0.0")
• dataset_id (string, optional): Dataset used for training
• tags (array, optional): Organizational tags
• description (string, optional): Model description

Example:

{
  "model_id": "sentiment_v1",
  "version": "1.0.0",
  "dataset_id": "reviews_2025",
  "tags": ["production", "sentiment"],
  "description": "Production sentiment classifier"
}

load_model_persistent

Load a model from SurrealDB into memory.

Parameters:

• model_id (string, required): Model to load
• version (string, optional): Specific version (defaults to latest)

list_model_versions

List all persisted versions of a model.

Parameters:

• model_id (string, required): Model to list versions for

store_training_dataset

Save training data to SurrealDB for reproducibility.

Parameters:

• dataset_id (string, required): Unique dataset identifier
• training_data (array, required): Array of {text, label} objects
• metadata (object, optional): Dataset metadata

Example:

{
  "dataset_id": "reviews_2025",
  "training_data": [
    {"text": "Great product!", "label": "positive"},
    {"text": "Disappointed", "label": "negative"}
  ],
  "metadata": {
    "source": "customer_reviews",
    "date": "2025-01-15"
  }
}

load_training_dataset

Load a stored training dataset from SurrealDB.

Parameters:

• dataset_id (string, required): Dataset to load

Monitoring Tools

get_model_metrics

Get performance metrics for a model over time.

Parameters:

• model_id (string, required): Model to analyze
• time_range (object, optional): Time window
  • start (string): ISO datetime
  • end (string): ISO datetime

Returns:

• accuracy: Overall accuracy (if ground truth provided)
• total_predictions: Number of predictions
• avg_confidence: Average confidence score
• avg_latency_ms: Average inference time
• predictions_per_label: Distribution by category

get_confusion_matrix

Get confusion matrix showing prediction accuracy by class.

Parameters:

• model_id (string, required): Model to analyze
• time_range (object, optional): Time window

Returns: Confusion matrix with true labels vs predicted labels.

detect_drift

Detect if model performance has drifted over time.

Parameters:

• model_id (string, required): Model to analyze
• baseline_window (object, required): Baseline time period
  • start (string): ISO datetime
  • end (string): ISO datetime
• current_window (object, required): Current time period
  • start (string): ISO datetime
  • end (string): ISO datetime

Returns:

• drift_detected: Boolean indicating significant drift
• drift_score: KL divergence score
• baseline_distribution: Label distribution in baseline
• current_distribution: Label distribution currently

Embedding Storage

store_embeddings

Store embeddings for similarity search.

Parameters:

• collection_name (string, required): Collection identifier
• items (array, required): Items with embeddings
  • item_id (string): Unique item ID
  • text (string): Original text
  • embedding (array): Vector embedding
  • metadata (object, optional): Additional data

search_similar

Search for similar items using cosine similarity.

Parameters:

• collection_name (string, required): Collection to search
• query_embedding (array, required): Query vector
• top_k (number, optional): Number of results (default: 5)

Usage Examples

Basic Text Classification

// Train a model
{
  "model_id": "sentiment",
  "training_data": [
    {"text": "Great product!", "label": "positive"},
    {"text": "Works perfectly", "label": "positive"},
    {"text": "Waste of money", "label": "negative"},
    {"text": "Very disappointing", "label": "negative"}
  ]
}

// Make predictions
{
  "model_id": "sentiment",
  "text": "This is amazing!",
  "top_k": 2
}

Production Workflow with Persistence

// 1. Store your dataset
{
  "dataset_id": "production_reviews_v1",
  "training_data": [...],
  "metadata": {"source": "app", "date": "2025-01-15"}
}

// 2. Train and immediately persist
{
  "model_id": "sentiment_prod",
  "training_data": [...],
  "persist": true,
  "version": "1.0.0",
  "dataset_id": "production_reviews_v1",
  "tags": ["production"]
}

// 3. Later, after server restart
{
  "model_id": "sentiment_prod"
}
// Model loads automatically from database!

// 4. Make predictions with monitoring
{
  "model_id": "sentiment_prod",
  "text": "This is great!",
  "log_prediction": true,
  "ground_truth": "positive"
}

// 5. Check performance metrics
{
  "model_id": "sentiment_prod"
}

Monitoring Model Performance

// Get metrics over last week
{
  "model_id": "sentiment_prod",
  "time_range": {
    "start": "2025-01-08T00:00:00Z",
    "end": "2025-01-15T00:00:00Z"
  }
}

// Get confusion matrix
{
  "model_id": "sentiment_prod"
}

// Detect drift (compare this week vs last week)
{
  "model_id": "sentiment_prod",
  "baseline_window": {
    "start": "2025-01-01T00:00:00Z",
    "end": "2025-01-08T00:00:00Z"
  },
  "current_window": {
    "start": "2025-01-08T00:00:00Z",
    "end": "2025-01-15T00:00:00Z"
  }
}

Semantic Search with Embeddings

// 1. Generate embeddings
{
  "model_id": "sentiment",
  "text": "Great customer service"
}

// 2. Store embeddings
{
  "collection_name": "reviews",
  "items": [
    {
      "item_id": "review_1",
      "text": "Great customer service",
      "embedding": [0.123, 0.456, ...],
      "metadata": {"rating": 5}
    }
  ]
}

// 3. Search similar
{
  "collection_name": "reviews",
  "query_embedding": [0.122, 0.458, ...],
  "top_k": 5
}

Architecture

astermind-elm-mcp/
├── src/
│   ├── index.ts              # Main MCP server with all tools
│   ├── model-manager.ts      # Model lifecycle management
│   ├── types.ts              # TypeScript type definitions
│   ├── persistence/
│   │   ├── surrealdb-client.ts  # Database client
│   │   └── types.ts             # Persistence types
│   └── scripts/
│       └── init-db.ts        # Database schema initialization
├── build/                    # Compiled JavaScript
├── package.json
├── tsconfig.json
└── README.md

Database Schema

The server uses SurrealDB with the following tables:

• models: Model storage with versioning
• datasets: Training data management
• predictions: Prediction logging for monitoring
• embeddings: Vector storage for similarity search

Schema is automatically created by running:

node build/scripts/init-db.js

Performance

• Training Time: Milliseconds for hundreds of examples
• Inference Time: Microseconds per prediction
• Memory: Efficient RAM usage
• Storage: SurrealDB persistence with ~1-5ms overhead
• No GPU Required: CPU-only, optimized performance

Technical Details

Extreme Learning Machines (ELM)

• Random hidden layer with fixed weights
• Closed-form output layer computation
• No iterative training (no gradient descent)
• Fast, efficient, and accurate

Persistence Layer

• SurrealDB for production-grade storage
• Models survive server restarts
• Dataset versioning and lineage tracking
• Prediction logging for monitoring
• Drift detection with KL divergence

Features

• Multiple activation functions (ReLU, LeakyReLU, Sigmoid, Tanh, Linear, GELU)
• Various weight initialization (Uniform, Xavier, He)
• Ridge regularization for stability
• Dropout support
• Character and token-based encoding

Troubleshooting

Server won't start

• Ensure dependencies: npm install
• Build TypeScript: npm run build
• Check Node.js version: Requires Node.js 18+

SurrealDB connection issues

• Verify SurrealDB is running: pgrep -f surreal
• Check connection URL matches config
• Initialize schema: node build/scripts/init-db.js

Models not persisting

• Ensure ENABLE_PERSISTENCE=true in environment
• Verify SurrealDB credentials
• Check database logs for errors

Poor predictions

• Increase hiddenUnits (try 256 or 512)
• Add more training examples
• Try different activation functions
• Adjust ridgeLambda for regularization

Dataset array issues

• Schema uses proper .* wildcard notation for nested fields
• Run node build/scripts/init-db.js to update schema
• Verify SurrealDB version (1.0.0+)

Development

Build

npm run build

Watch mode

npm run watch

Initialize database schema

node build/scripts/init-db.js

License

MIT

Credits

Built on AsterMind-ELM by infiniteCrank.

Related Documentation

• Quick Start Guide - Get started in 5 minutes
• Usage Guide - Detailed usage examples
• Persistence Guide - Deep dive into persistence features
• Quick Reference - Command reference