DataOps MCP Agent

MCP-powered agent integrating with Snowflake for autonomous data analysis.

Overview

This project showcases how to build production-ready AI agents that can:

• Query enterprise data warehouses using natural language
• Execute cross-database joins autonomously
• Maintain security through read-only access patterns
• Provide audit trails for compliance

Architecture

User Query
    ↓
Claude (Anthropic API)
    ↓ (MCP Protocol)
MCP Server (Python)
    ↓
Snowflake Tools Layer
    ↓
Snowflake Data Warehouse

Key Components

1. MCP Server: Standardized protocol for AI-tool integration
2. Snowflake Tools: Secure, read-only database query layer
3. Claude Integration: LLM orchestration with autonomous discovery
4. Cross-Database Queries: Enterprise pattern for federated data access

Technical Stack

• LLM: Claude 3.5 Sonnet (Anthropic)
• Protocol: Model Context Protocol (MCP)
• Data Platform: Snowflake
• Language: Python 3.10
• Key Libraries: anthropic, mcp, snowflake-connector-python

Business Value

For Enterprise Customers

• Secure by design: Read-only queries, no data exfiltration
• Governed access: All queries logged and auditable
• Cross-platform: MCP works with any compatible LLM
• Incremental adoption: Start with analytics, expand to operations

Features

• Natural language to SQL translation
• Cross-database join execution
• Autonomous schema discovery
• Business-focused response formatting
• Query safety validation (SELECT-only)
• Conversation context retention

Quick Start

Prerequisites

• Python 3.10+
• Snowflake trial account (free at snowflake.com)
• Anthropic API key

Installation

git clone https://github.com/pushkersahai/dataops-mcp-agent.git
cd dataops-mcp-agent

python -m venv venv
source venv/Scripts/activate  # Windows Git Bash
# source venv/bin/activate     # Mac/Linux

pip install -r requirements.txt

cp .env.example .env
# Edit .env with your credentials

Setup Demo Data

python src/setup_demo_data.py

Creates two databases:

• CUSTOMER_DB.MASTER.CUSTOMERS (10 customers)
• SALES_DB.TRANSACTIONS.SALES_ORDERS (15 orders)

Run the Agent

# Autonomous mode (recommended for demos)
python src/claude_mcp_client_v2.py

# Interactive mode (asks for clarification)
python src/claude_mcp_client.py

Example Usage

You: What are the top 5 largest closed deals?

Claude is using 1 tool(s):
  - execute_query

Claude: Based on the sales data, here are the top 5 largest closed deals:

1. Retail Giants Inc - Snowflake Enterprise License: $1,200,000
2. Finance Partners Group - Snowflake Enterprise License: $1,100,000
3. Energy Solutions Corp - Snowflake Enterprise License: $950,000
4. Finance Partners Group - AWS Data Lake Solution: $890,000
5. Acme Corporation - Snowflake Enterprise License: $450,000

Total value: $4,590,000

More examples:

• "Show me total revenue by industry"
• "Which account manager has the most closed revenue?"
• "What's our average deal size for Enterprise customers?"
• "Show me all Technology customers with their purchase history"

Project Structure

dataops-mcp-agent/
├── src/
│   ├── mcp_server/
│   │   └── server.py                 # MCP server implementation
│   ├── tools/
│   │   └── snowflake_tools.py        # Snowflake query tools
│   ├── claude_mcp_client.py          # Interactive mode
│   ├── claude_mcp_client_v2.py       # Autonomous mode
│   ├── setup_demo_data.py            # Demo data generator
│   ├── test_snowflake_connection.py  # Connection validator
│   └── test_mcp_server.py            # Tool testing
├── docs/
│   ├── ARCHITECTURE.md               # Technical deep dive
│   └── PRODUCTIONIZATION.md          # Deployment guide
├── .env.example                      # Environment template
├── requirements.txt                  # Dependencies
└── README.md                         # This file

Security & Governance

Built-in Safeguards

1. Read-only access: Only SELECT queries permitted
2. Query validation: Blocks DROP, DELETE, UPDATE, INSERT
3. No data storage: Results streamed, not persisted
4. Credential isolation: API keys in environment variables
5. Audit trail: All queries logged with timestamps

Enterprise Considerations

• Role-Based Access Control (RBAC) in Snowflake
• Network policies for IP whitelisting
• OAuth integration for SSO
• Query result size limits
• Rate limiting on API calls

How MCP Differs from Traditional RAG

Aspect	Traditional RAG	MCP Agent
Data Access	Copy to vector DB	Query live data
Freshness	Stale (periodic sync)	Real-time
Security	Data duplication	Source of truth
Maintenance	Embedding updates	Schema-aware
Cost	Storage + compute	Query compute only

Production Deployment

See docs/PRODUCTIONIZATION.md for:

• AWS Lambda deployment
• Container orchestration (ECS/Kubernetes)
• Monitoring and observability
• Cost optimization strategies
• Multi-tenant architectures
• Compliance and governance

Testing

# Test Snowflake connection
python src/test_snowflake_connection.py

# Test MCP tools locally
python src/test_mcp_server.py

Roadmap

Potential enhancements:

• Slack integration for team queries
• dbt Cloud API integration for lineage analysis
• Web UI with Streamlit
• Query result caching
• Natural language to dbt model generation
• Multi-LLM support (GPT-4, Gemini)

Related Projects

• Macro Regime Decision Agent - dbt + Snowflake financial analysis

Why MCP?

Model Context Protocol standardizes AI-tool integration:

1. Interoperability: Same tools work with Claude, GPT-4, Gemini
2. Reusability: Write once, use across projects
3. Governance: Centralized access control
4. Future-proof: Swap LLMs without rewriting integrations

License

MIT License

Author

Pushker Sahai www.linkedin.com/in/pushkersahai