Windsor MCP Server: SQL-Free LLM Data Access

Overview

Windsor MCP Server provides an interface between large language models (LLMs) and your full-stack business data without requiring SQL or custom scripting. It implements a Model Context Protocol (MCP) server in the Windsor.ai stack that publishes schema-aware, queryable context about your connected data sources so an LLM can reason about and answer business questions using structured data semantics.

For developer teams, the server removes the need to hand-craft SQL or build ad-hoc APIs for each data source. Instead, Windsor MCP exposes a consistent, documented surface that describes tables, fields, sample rows and precomputed summaries. LLMs (or applications using LLMs) can call the MCP server to fetch contextual data, run analytical queries, and receive structured results suitable for report generation, automated insights, or downstream visualization.

Features

• SQL-free access: Ask business questions in natural language; the MCP layer maps intent to data context instead of requiring raw SQL.
• Schema-first context: Exposes table/column metadata, sample rows and relationships so models understand your data layout.
• Connectors-managed: Works with Windsor.ai connectors to surface multiple data sources (analytics, advertising, CRM, BI).
• Tooling for LLMs: Provides tools and endpoints intended to be invoked by models or orchestrators to fetch relevant context.
• Access control and auditing: Integrates with Windsor authentication to control who/what can request data.
• Extensible: Add custom data sources or precomputed metrics for domain-specific workflows.

Installation / Configuration

Below are typical steps to get the MCP server running locally or in a containerized environment. Refer to the repository for exact commands and any platform-specific details.

Clone the repo and inspect the README:

git clone https://github.com/windsor-ai/windsor_mcp.git
cd windsor_mcp

Build and run with Docker (example):

# Build (if repository contains a Dockerfile)
docker build -t windsor_mcp .

# Run, exposing a local port and passing environment variables
docker run -p 8080:8080 \
  -e WINDSOR_API_KEY="your_api_key_here" \
  -e MCP_ENV="production" \
  windsor_mcp

Or run locally (example for a Python/Node service—check repo for exact requirements):

# Python example (if applicable)
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
python -m windsor_mcp.app --port 8080

Configuration is typically provided via environment variables or a JSON/YAML config file. Example JSON config snippet for data sources:

{
  "data_sources": [
    {
      "name": "ga4",
      "type": "google_analytics",
      "credentials": "/path/to/credentials.json"
    },
    {
      "name": "ads",
      "type": "google_ads",
      "credentials": { "client_id": "...", "client_secret": "..." }
    }
  ],
  "mcp": {
    "max_context_rows": 100,
    "sample_rows_per_table": 10
  }
}

Common environment variables (example):

Always consult the repo’s documentation for exact names and supported connectors.

Available Resources

• GitHub repository: https://github.com/windsor-ai/windsor_mcp — source, issues, and changelog
• Example configs and connector definitions (in the repo’s examples/ directory)
• API docs and OpenAPI spec (if provided in the repository) — useful for generating client code
• Windsor.ai connector catalog — lists supported integrations (analytics, ads, CRM, databases)

Typical API Interaction

The MCP server is generally used by an orchestrator or LLM agent. A common pattern:

1. Agent asks the MCP server for schema/context relevant to a natural-language question.
2. MCP responds with table descriptions, sample rows, and precomputed summaries.
3. Agent requests specific subsets of data or metrics; MCP returns structured JSON.

Example request (generic):

curl -X POST http://localhost:8080/mcp/context \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "question": "Show last quarter revenue by marketing channel",
    "max_rows": 50,
    "tables": ["revenue", "campaigns"]
  }'

Example response (conceptual):

{
  "schema": {
    "revenue": { "columns": ["date", "channel", "revenue"] },
    "campaigns": { "columns": ["campaign_id", "name", "channel"] }
  },
  "samples": { "revenue": [ /* rows */ ], "campaigns": [ /* rows */ ] },
  "metrics": { "revenue_by_channel": [ {"channel": "Paid", "revenue": 12345}, ...] }
}

Check the repository docs for exact endpoints and payload formats.

Use Cases

• Ad-hoc analytics without SQL: Business users or LLM agents ask for metrics like “Top 5 campaigns by ROAS in the past 30 days” and receive structured metric outputs and context to validate.
• Automated reporting: