Snowflake Cortex MCP Server for AI Features

Overview

Snowflake Cortex MCP Server is an implementation of the Model Context Protocol (MCP) that bridges MCP-compatible clients (for example: Claude for Desktop, fast-agent, or the Agentic Orchestration Framework) with data stored and managed inside Snowflake. It exposes HTTP endpoints that MCP clients use to request contextual data, embeddings, and other retrieval primitives so that LLMs and autonomous agents can operate on up-to-date enterprise data without embedding credentials or data directly into the client.

This server is useful when you want to enable AI features that require retrieval from Snowflake (e.g., retrieval-augmented generation, tool execution, or knowledge-grounding) while keeping Snowflake credentials and business logic centralized. The repository includes a lightweight MCP-compatible adapter that translates MCP calls into Snowflake queries and returns structured context to the client.

Features

• MCP-compatible HTTP endpoints for context, embeddings, and model metadata
• Secure Snowflake connectivity (configurable credentials and roles)
• Optional token-based authentication for MCP clients
• Simple Docker and local development options
• Designed to integrate with MCP clients such as:
  • Claude for Desktop
  • fast-agent
  • Agentic Orchestration Framework (AOF)
• Extensible: plug-in your own SQL or retrieval logic to control how contextual data is composed

Installation / Configuration

Prerequisites:

• A Snowflake account with a user/role that can run the queries you need
• Docker (recommended) or a suitable runtime (Node/Python) depending on the repository implementation

1. Clone the repository:

git clone https://github.com/thisisbhanuj/Snowflake-Cortex-MCP-Server.git
cd Snowflake-Cortex-MCP-Server

2. Configure environment variables. Example .env:

# MCP server
MCP_SERVER_HOST=0.0.0.0
MCP_SERVER_PORT=8000
MCP_AUTH_TOKEN=change_this_to_a_secret_token

# Snowflake connection
SNOWFLAKE_ACCOUNT=your_account_region
SNOWFLAKE_USER=your_user
SNOWFLAKE_PASSWORD=your_password
SNOWFLAKE_ROLE=ANALYST
SNOWFLAKE_DATABASE=MY_DATABASE
SNOWFLAKE_SCHEMA=PUBLIC
SNOWFLAKE_WAREHOUSE=COMPUTE_WH

3a) Run with Docker (recommended):

# Build (if Dockerfile present)
docker build -t snowflake-cortex-mcp-server .

# Run
docker run -d \
  -p 8000:8000 \
  --env-file .env \
  --name snowflake-mcp \
  snowflake-cortex-mcp-server

3b) Run locally (if project uses Node.js or Python; check repo instructions):

# Node.js example
npm install
npm start

# OR Python / FastAPI example
pip install -r requirements.txt
uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload

4. Verify health:

curl http://localhost:8000/health
# expected: 200 OK / JSON status

Notes:

• Inspect the repository README for exact run commands and dependencies (Node vs Python).
• Use a secure secret manager for production credentials and avoid committing .env files.

Available Resources

• GitHub repository: https://github.com/thisisbhanuj/Snowflake-Cortex-MCP-Server
• Typical MCP endpoints (implementation-dependent — check repo):
  • GET /health — service health
  • GET /mcp/v1/models — available models / metadata
  • POST /mcp/v1/context — retrieve contextual documents/rows
  • POST /mcp/v1/embeddings — (optional) produce embeddings or proxy to a model
• OpenAPI / Swagger: check the repo for an OpenAPI spec or route documentation.
• Example client configs: sample request snippets and token usage are usually provided in the repo.

Environment variable summary:

Use Cases

• Retrieval-Augmented Generation (RAG) for customer support:

  • A chat client (Claude for Desktop) sends a query via MCP to the server.
  • The server runs parameterized SQL on Snowflake to fetch relevant rows, formats them into context, and returns them to the LLM for grounded responses.

• Agent orchestration with fast-agent or Agentic Orchestration Framework:

  • Autonomous agents request structured facts or table rows through MCP endpoints.
  • The server enforces RBAC and auditability because all Snowflake access is centralized.

• Internal knowledge assistant on the desktop:

  • Developers or analysts run Claude for Desktop configured to use this MCP server URL and token. The assistant can reference internal data (product catalogs,