Docs / Desktop AI access (MCP)
Desktop AI access (MCP)
Let an AI agent discover connections, describe tables and run paginated queries through a loopback MCP server.
The MCP title-bar button starts an authenticated, loopback-only Model Context Protocol server. It exposes saved connection discovery, connection opening, table metadata and paginated queries through the same adapters the interface uses. This feature is desktop-only.
Tools
| Tool | Purpose |
|---|---|
list_databases | List saved connections without stored passwords |
connect | Open a saved connection |
list_tables | List tables and views in a schema |
describe_table | Return column metadata for a table |
run_query | Run a paginated query with backend limits |
Security
- The server binds to loopback, never the Wi-Fi interface, and requires an unpredictable bearer token. Host and origin restrictions are enforced.
- Enablement is explicit and session-scoped. The panel shows concrete client configuration and server status.
- MCP respects the connection's No-Edit and read-only policy and cannot unlock a connection.
- Query timeouts, pagination and maximum page sizes match backend limits. Errors and cancellation propagate to the caller.
Client configuration
Copy the endpoint and token from the MCP panel into your agent's MCP client settings. The token changes each time the server starts. A typical configuration entry looks like this:
{
"mcpServers": {
"querybench": {
"url": "http://127.0.0.1:<port>/mcp",
"headers": { "Authorization": "Bearer <token from the MCP panel>" }
}
}
}