The Model Context Protocol changed how AI tools see your data. Here's what's actually happening, in plain English.
MCP (Model Context Protocol) is an open standard that lets AI tools like Claude read structured context (files, databases, APIs) without custom integration. An MCP server for your data warehouse exposes your schema, gotchas, metric definitions, and query capability over MCP, so any AI client can see them. That's how Contextary connects your warehouse to Claude in 60 seconds.
Before MCP, getting AI to work with your data warehouse meant pasting your schema into a prompt, hoping it survived the context window, and re-pasting it the next day. Every team built a custom integration. Every tool spoke a different dialect. Every prompt drifted.
MCP is the equivalent of HTTP for AI context. One open protocol that any AI client can speak, against any server that exposes context. Build the server once, plug it into Claude, Cursor, Claude Code, or anything else that supports MCP. No custom integration per tool.
The AI tool (Claude Desktop, Cursor, Claude Code) is the client. The thing that holds your context is the server. MCP defines the conversation between them.
When you connect a new MCP server, the AI client asks "what do you offer?" and gets back the catalog automatically. You don't have to teach Claude what your warehouse looks like. The server does.
When the user asks a question, the AI fetches just the relevant context (say, the opportunities table annotations and the "qualified pipeline" metric definition) instead of stuffing everything into the prompt up front. That's why MCP scales to warehouses with hundreds of tables; it loads context on demand.
Contextary runs a hosted MCP server for your warehouse. Once you connect Snowflake, BigQuery, Postgres, or Redshift and annotate your handbook, the server makes the following available to any MCP-speaking AI client:
Schema metadata only, never raw data, leaves your warehouse to reach the AI provider. That's the security model. Read the full security overview for the details.
You have two options:
Sign up for Contextary, connect your warehouse, and Contextary runs the MCP server for you at https://app.contextary.ai/mcp. Drop one line into Claude's config and you're done:
{
"mcpServers": {
"contextary": {
"url": "https://app.contextary.ai/mcp"
}
}
}Anthropic's MCP SDK is open-source. You can write a Python or TypeScript server that connects to your warehouse, manages your annotations, handles auth, and ships verification, and operate it yourself. Reasonable choice for orgs that want full control. Most teams find the cost of building and maintaining the MCP layer doesn't pay back when an off-the-shelf server already covers the case.
The four layers of context the MCP server delivers, and how to capture them.
Why an MCP server with real context matters more than longer prompts.
The metric-definition layer of an MCP server, in practical terms.
The setup flow, end to end. Same shape for BigQuery, Postgres, Redshift.
One MCP URL, every AI tool. Contextary runs the server, you keep the context. Free to start.