To set up NVIDIA Auto Ontology MCP for Cursor or Claude Desktop, first prepare an Auto Ontology deployment with a live database connection and a compiled semantic layer. Then run the MCP server and configure the client: Cursor connects directly to its /mcp URL, while Claude Desktop uses mcp-remote through npx. NVIDIA’s Auto Ontology repository and MCP documentation provide the setup details.
What you need before connecting a client
The MCP server is an HTTP client for Auto Ontology’s public API. It does not hold database credentials or model configuration, and its documented tools are read-only. A successful data query also requires a live database connection and a compiled semantic layer; having the MCP endpoint running is not enough.
For NVIDIA’s sample setup, have Docker with Compose v2, uv, and an NVIDIA API key ready. The quickstart uses Microsoft’s Wide World Importers sample database. You’ll also need the environment values for the model API key, AUTH_SECRET, APP_URL, POSTGRES_PASSWORD, CONNECTION_STRINGS, AUTO_ONTOLOGY_ADMIN_PASSWORD, and PGADMIN_PASSWORD.
Prepare Auto Ontology and start the MCP server
Follow these steps in order. The MCP server connects to the running Auto Ontology deployment, so prepare that deployment before configuring a client.
- Clone NVIDIA’s Auto Ontology repository. The quickstart’s environment file and setup instructions are in the project repository.
- Create and configure
.env. Copy.env.exampleto.env, then set the API key,AUTH_SECRET,APP_URL, database connection and password, administrator password, and pgAdmin password. The sample includes per-model API-key entries; replace them or remove them to useDEFAULT_MODELS_API_KEYas the fallback. - Replace the default passwords. Compose publishes PostgreSQL, pgAdmin, and the web app on the host, so set your own passwords before starting the services.
- Start PostgreSQL and load the sample database. NVIDIA’s quickstart uses its seed script for Wide World Importers.
- Start the rest of the Compose stack. This brings up the Auto Ontology web app and its supporting services.
- Enable semantic compilation in the web app. Sign in with the administrator account, open Settings → Semantic Compilation, and wait for the “Semantic compilation is ready” status.
- Launch the MCP package against the web app. NVIDIA’s instructions use
uvxwith the package in the repository’smcpsubdirectory andAUTO_ONTOLOGY_API_URL=http://localhost:3000.
The client-facing endpoint is http://localhost:3003/mcp. Keep the /mcp suffix in the URL.
Configure Cursor or Claude Desktop
The clients use the same endpoint but different configuration methods. Cursor registers it as a direct HTTP URL; NVIDIA’s Claude Desktop example starts npx with mcp-remote.
| Client | Configuration method | Endpoint setting |
| Cursor | In ~/.cursor/mcp.json, set mcpServers.auto-ontology.url. | http://localhost:3003/mcp |
| Claude Desktop | In claude_desktop_config.json, set command to npx and args to mcp-remote and the endpoint. | http://localhost:3003/mcp |
After saving the client configuration, restart Cursor or Claude Desktop and complete the Auto Ontology sign-in when prompted. Keep credentials out of the MCP client configuration: the client handles the login flow and its token, and each call uses the signed-in user’s permissions.
Diagnose sign-in and query problems
Check the deployment before changing a question or client setting. NVIDIA’s MCP server provides nine read-only tools, including check_readiness, check_answerable, ask_question, and search_terms.
- If sign-in discovery fails, check the deployment’s
/.well-known/oauth-authorization-serverroute. The documented sign-in flow expects an HTTP 200 response with JSON; an older prebuilt image may not expose the required route. - If the protected-resource URL does not match, check whether the client uses
localhostwhile the deployment is configured with127.0.0.1, or vice versa. SetAUTO_ONTOLOGY_MCP_PUBLIC_URLto match the address the client actually uses. - If a question returns no useful answer, run
check_readinessfirst. It identifies blockers such as missing semantic compilation or a missing live database connection. A database catalog can remain visible after its original connection is gone, so catalog entries alone do not establish that a query can run. - If readiness passes but the question still fails, use
check_answerableto check semantic coverage, thensearch_termsto find the vocabulary Auto Ontology recognizes. Once the question is answerable, useask_question; it returns an answer, SQL, and rows, with results capped at 100 rows. - If a viewer gets a 403 when reading connection settings, that setting can be admin-only; the error does not by itself indicate a failed deployment.
Start with check_readiness: it helps separate a deployment problem from a question-vocabulary problem.
Keep backend and ingestion services private
NVIDIA directs operators to keep backend port 3001 and ingestion port 3002 on a private network. The backend trusts identity forwarded by the frontend, and ingestion operations have no independent authentication. The MCP server itself does not hold database credentials; calls use the permissions of the user who signed in.