MCP Quoting System

An MCP (Model Context Protocol) based intelligent quoting system that compares incoming RFPs against historical quotes to generate accurate cost estimates and lead times.

🚀 Quick Start (Windows)

New to the system? Just double-click START.bat and choose option [1] for automatic setup!

See these guides:

• QUICKSTART.md - Visual guide with workflows and examples

• BATCH-FILES-README.md - Complete batch file documentation

• TEST-CASES.md - Sample RFQs to test the system

Available Batch Files

• START.bat - Main interactive launcher (recommended) ⭐

• setup.bat - First-time installation wizard

• start-dev.bat - Start development server

• start-prod.bat - Start production server

• quick-test.bat - Automated testing

• stop.bat - Stop the server

• See BATCH-FILES-README.md for complete list

Features

• RFP Parsing: Automatically extracts material, processes, quantities, tolerances, and other key information from text-based RFPs

• Historical Matching: Compares new requests against past quotes using intelligent similarity scoring

• Cost Estimation: Activity-based costing with material, processing, labor, tooling, and overhead calculations

• Lead Time Prediction: Estimates delivery time based on quantity, processes, and historical data

• Confidence Scoring: Provides low/medium/high confidence ratings based on data completeness and match quality

• Human-in-Loop: Requires approval before sending quotes, with full audit trails

• Idempotency: Prevents duplicate processing of the same RFP

Architecture

The system is built as an MCP server with the following capabilities:

MCP Functions (Capabilities)

1. ingestRfp - Parse RFP text and extract structured information

2. findSimilarQuotes - Search historical database for similar past quotes

3. estimateCostLeadTime - Calculate cost and lead time estimates

4. generateQuote - Create formatted quote documents

5. approveQuote - Mark quotes as approved (human-in-loop)

6. sendQuote - Send quotes via email (dry-run enabled)

Coordinator

• evaluateRfpAndDraftQuote - Orchestrates all functions to produce a complete quote evaluation

Installation

npm install

Configuration

1. Copy .env.example to .env:

cp .env.example .env

2. Edit .env with your settings:

PORT=3789
SMTP_HOST=smtp.gmail.com
SMTP_PORT=587
SMTP_USER=your-email@example.com
SMTP_PASS=your-app-password

Usage

Start the Server

Development mode:

npm run dev

Production mode:

npm run build
npm start

Load Sample Historical Data

Copy sample quotes to the main database:

cp data/sample-quotes.json data/quotes.json

Example API Calls

1. Full Quote Evaluation (Coordinator)

curl -X POST http://localhost:3789/mcp/invoke/evaluateRfpAndDraftQuote \
  -H "Content-Type: application/json" \
  -d '{
    "rfp": {
      "rawText": "We need 200 pcs of a 6061-T6 aluminum widget, CNC machined, anodize finish, tolerance +/-0.005, delivery by 2025-02-28. Contact: buyer@acme.com",
      "qty": 200,
      "contactEmail": "buyer@acme.com",
      "customerName": "Acme Corp"
    }
  }'

2. Get Formatted Review

curl -X POST http://localhost:3789/mcp/utility/formatReview \
  -H "Content-Type: application/json" \
  -d '{
    "result": {<evaluation_result_from_previous_call>}
  }'

3. View Historical Quotes

curl http://localhost:3789/mcp/utility/historicalQuotes

4. Add Historical Quote

curl -X POST http://localhost:3789/mcp/utility/addHistoricalQuote \
  -H "Content-Type: application/json" \
  -d '{
    "id": "Q-NEW",
    "quoteDate": "2024-11-12T10:00:00Z",
    "customerName": "New Customer",
    "normalized": {
      "material": "steel",
      "processes": ["laser", "bend"],
      "qtyRange": [51, 100],
      "tolerances": "+/-0.010"
    },
    "costPerUnit": 25.00,
    "totalCost": 1875.00,
    "leadDays": 14,
    "approved": true
  }'

Similarity Matching

The system uses rule-based similarity scoring with weighted components:

• Material (35%): Exact, family, or partial matches

• Processes (30%): Overlap of required processes

• Quantity (20%): Same range or adjacent ranges

• Tolerances (10%): Matching precision requirements

• Finish (5%): Surface treatment matching

Confidence Thresholds

• High confidence (≥85%): Very similar to past work, reliable estimate

• Medium confidence (70-85%): Similar family, adjust with caution

• Low confidence (<70%): New type of work, requires engineer review

Cost Estimation

Activity-based costing model:

Total Cost = Material + Processing + Labor + Tooling + Overhead + Margin

Components

• Material Cost: Unit price × quantity (from material price list)

• Processing Cost: Sum of process times × machine hour rate

• Labor Cost: Operator time × labor rate

• Tooling Amortization: Setup cost / quantity

• Overhead: 15% of direct costs

• Margin: 20% profit margin

• Contingency: 10% for low-confidence quotes

Lead Time Calculation

Lead Time = Procurement + Setup + Run Time + QA + Shipping

Adjustments based on:

• Quantity (higher volume = longer lead time)

• Process complexity (heat treat, plating add time)

• Historical actual lead times from similar quotes

Data Storage

Currently uses JSON files in the data/ directory:

• quotes.json - Historical quotes database

• evaluations.json - Recent RFP evaluations (last 100)

For production, consider migrating to:

• PostgreSQL for relational data

• Vector database (Pinecone, Weaviate) for semantic similarity search

• Redis for caching and idempotency

Safety Features

1. Human-in-Loop: All quotes default to "draft" status

2. Dry-Run Email: Email sending requires explicit enablement

3. Idempotency: Duplicate RFPs return cached results

4. Audit Trails: All evaluations logged with timestamps

5. Confidence Scoring: Flags uncertain estimates for review

Extending the System

Add New Materials

Edit src/config.ts:

materials: {
  'titanium-grade-5': 18.0,
  // Add more...
}

Add New Processes

Edit src/config.ts:

processes: {
  'EDM': 40,  // minutes per part
  'Grinding': 25,
  // Add more...
}

Integrate Vector Search

Replace the rule-based matcher in src/matcher.ts with:

1. OpenAI embeddings for RFP text

2. Vector DB (Pinecone, Weaviate, FAISS)

3. Cosine similarity search

4. Metadata filtering (material, process)

Add Database Backend

Replace src/storage.ts with database adapters:

• Use Prisma or TypeORM for PostgreSQL

• Implement connection pooling

• Add transactions for data integrity

Testing

Create test RFPs:

// Test 1: High similarity match
{
  "rawText": "Need 250 units of 6061-T6 aluminum, CNC milled and anodized, +/-0.005 tolerance",
  "qty": 250
}

// Test 2: New material
{
  "rawText": "100 titanium brackets, laser cut and polished",
  "qty": 100
}

// Test 3: Low detail (low confidence)
{
  "rawText": "We need some metal parts",
  "qty": 50
}

API Documentation

See full API documentation in the console output when starting the server.

Troubleshooting

No historical matches found

• Check that data/quotes.json exists and has content

• Verify material names match (case-insensitive)

• Lower similarity threshold in src/config.ts

Costs seem incorrect

• Review material prices in src/config.ts

• Adjust machine hour rate and labor rate

• Check overhead and margin percentages

Lead times too short/long

• Adjust defaultLeadDays in config

• Review process time estimates

• Check quantity-based scaling logic

Future Enhancements

1. ML-based similarity: Train model on historical quote-to-win patterns

2. Drawing analysis: Extract features from CAD/PDF drawings

3. Supplier integration: Real-time material lead times from vendors

4. CRM integration: Auto-populate customer info

5. Dashboard UI: React frontend for engineers to review/approve

6. Analytics: Win/loss tracking, pricing optimization

7. Multi-currency: International quote support

8. Revision tracking: Quote version history

License

MIT

Support

For issues or questions, please contact your system administrator.