Skip to main content

Databricks Agent Bricks

Build a deployed agent on Databricks that combines SQL execution (via a Genie Space) with DataHub's catalog context in a single conversation. The agent discovers tools from both MCP endpoints and the LLM decides which to call per request — search DataHub to find the right table, then query the Genie Space for the actual data.

Prerequisites​

Create a Genie Space​

You need a Genie Space with the tables you want the agent to query. If you don't have one yet, create one in the Databricks UI under Genie. Grab the Space ID from the URL: https://<workspace>.databricks.com/genie/rooms/<space-id>.

Build the Agent​

Full example: agent.py

The core idea — connect to both MCP endpoints and build a unified tool registry:

def _collect_mcp_tools(workspace_client: WorkspaceClient):
    host = workspace_client.config.host
    server_urls = [
        f"{host}/api/2.0/mcp/genie/{os.getenv('GENIE_SPACE_ID')}",
        f"{host}/api/2.0/mcp/external/{os.getenv('DATAHUB_CONNECTION_NAME', 'datahub')}",
    ]
    all_tools, clients = [], {}
    for url in server_urls:
        client = DatabricksMCPClient(server_url=url, workspace_client=workspace_client)
        for tool in client.list_tools():
            all_tools.append(tool)
            clients[tool.name] = client
    return all_tools, clients

Run It​

export GENIE_SPACE_ID="your-genie-space-id"          # from the Genie Space URL (see above)
export DATAHUB_CONNECTION_NAME="datahub"              # name of your UC connection
export OPENAI_API_KEY="sk-..."                        # or use Databricks Foundation Models

databricks auth login --host https://your-workspace.cloud.databricks.com
uv run start-app
# Chat at http://localhost:8000

For deployment, see Databricks' Apps deployment guide. You can also swap OpenAI for a Databricks Foundation Model by pointing the OpenAI client at your workspace's serving endpoint.

Example​

User: What are our top customer segments? Who owns that data?

Agent: (searches DataHub, then queries the Genie Space)

The customer_segments table is owned by the Marketing Analytics team. Top segments: High-Value Loyalists (45k), At-Risk Champions (32k), New Engaged (28k).

For a full list of available DataHub tools, see the Agent Context Kit.

Troubleshooting​

  • DataHub not showing up? Check the connection under Catalog > Connections, verify the MCP URL and base path, and confirm the bearer token is still valid.
  • 401 Unauthorized? Regenerate the DataHub access token and update the UC connection.
  • No tools discovered? Make sure DataHub's MCP server is enabled. Base path is /integrations/ai/mcp (Cloud) or /mcp (self-hosted).

Links: MCP on Databricks · DataHub MCP Server · Agent Context Kit