> ## Documentation Index
> Fetch the complete documentation index at: https://ventstream.dev/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# MCP over live data

> A public MCP endpoint an AI agent can query — over documents that update in real time.

VentStream's engine ships a read-only [MCP server](/docs/guides/mcp-server). This
demo exposes one publicly, over documents that live pipelines keep in sync
from Postgres and MySQL sources in our acceptance cluster.

## Point Claude at it

Add the endpoint to Claude Desktop, Claude Code, or any MCP client:

```json theme={"dark"}
{
  "mcpServers": {
    "ventstream-demo": {
      "type": "http",
      "url": "https://mcp-demo.ventstream.dev/mcp",
      "headers": {
        "Authorization": "Bearer vsk_1xscwJ4l0to5pD-tiFXMQuZ8AJBvDLtoG_73E85VBwc"
      }
    }
  }
}
```

That bearer token is a real, published demo key — read-only, scoped to the
demo dataset, and rotated periodically. Then ask things like:

* *"What targets can you see?"*
* *"Find the five most recent orders and summarize them."*
* *"Any orders updated in the last five minutes?"* — there will be; the
  generators never stop.

## Why it's interesting

The agent holds **no database credentials**. It reads the sink documents a
pipeline materializes — composed, current, and scoped by its token: a key
scoped to one target simply cannot see the others (out of scope reads as
nonexistent). That's the pattern for giving AI tools live operational data
without handing them your database.

The [MCP server guide](/docs/guides/mcp-server) covers running this against your
own pipelines, and `examples/mcp` in the repo is the full worked example.