Procedure

1. Configure Cursor as a client.

   1. From your Desktop app, navigate to Cursor Settings and select MCP Tools > New MCP Server.

   2. Add the following configuration:

      {
        "mcpServers": {
          "backstage-actions": {
            "url": "https://<my_developer_hub_domain>/api/mcp-actions/v1",
            "headers": {
              "Authorization": "Bearer <MCP_TOKEN>"
            }
          }
        }
      }

      where:

      <MCP_TOKEN>

      Enter the previously configured static token

      <my_developer_hub_domain>

      Enter the hostname of your RHDH instance

2. Configure Continue as a client.

   1. In your agent yaml configuration file, add the following configuration:

      mcpServers:
        - name: backstage-actions
          type: sse
          url: https://<my_developer_hub_domain>/api/mcp-actions/v1/sse
          requestOptions:
            headers:
              Authorization: "Bearer <MCP_TOKEN>"

      where:

      <MCP_TOKEN>

      Enter the previously configured static token

      <my_developer_hub_domain>

      Enter the hostname of your RHDH instance

3. Configure Developer Lightspeed for RHDH as a client. For more details, see Red Hat Developer Lightspeed for Red Hat Developer Hub.

   1. In the lightspeed-stack.yaml configuration, add the following configuration for mcp_servers:

      mcp_servers:
        - name: mcp::backstage
          provider_id: model-context-protocol
          url: https://<my_developer_hub_domain>/api/mcp-actions/v1
          authorization_headers:
            Authorization: "client"

      where:

      model-context-protocol

      This is the tool runtime provider defined and configured in the llama-stack run.yaml configuration for use in LCS.

   2. Optional: If you want to use your own Llama Stack configuration, add the following code to your Llama Stack configuration file (run.yaml).

      providers:
        tool_runtime:
        - provider_id: model-context-protocol
          provider_type: remote::model-context-protocol
          config: {}

   3. To authorize requests to the MCP endpoint using <MCP_TOKEN>, add it in the Developer Lightspeed for RHDH app-config.yaml file, to make POST requests to LCS, as shown in the following code:

      lightspeed:
        mcpServers:
        - name: mcp::backstage
          token: ${MCP_TOKEN}

   4. Optional: You can query the LCS /v1/streaming_query endpoint directly by providing the MCP_TOKEN in the header, as shown in the following code:

      curl -X POST \
        -H `Content-Type: application/json` \
        -H `MCP-HEADERS: {"mcp::backstage": {"Authorization": "Bearer <MCP_TOKEN>"}}` \
        -d `{"query": "Can you give me all catalog templates of type 'service', "model": "gpt-4o-mini", "provider": "openai"}` \
        _<url>_/v1/streaming_query

Procedure

1. Log in to the OCP cluster running RHDH and go to your RHDH project using the following code:

   $ oc project my-rhdh-project

2. Inspect the logs for the installation of the RHDH dynamic plugins using the following code:

   $ oc logs -c install-dynamic-plugins deployment/<my-product-deployment>

Verification

1. You must see an entry for the MCP backend server plugin as shown in the following code:

   ..... prior logs ....
   ======= Installing dynamic plugin oci://ghcr.io/redhat-developer/rhdh-plugin-export-overlays/backstage-plugin-mcp-actions-backend:bs_1.42.5__0.1.2!backstage-plugin-mcp-actions-backend
   	==> Copying image oci://ghcr.io/redhat-developer/rhdh-plugin-export-overlays/backstage-plugin-mcp-actions-backend:next__0.2.0 to local filesystem
   	==> Successfully installed dynamic plugin oci://ghcr.io/redhat-developer/rhdh-plugin-export-overlays/backstage-plugin-mcp-actions-backend:bs_1.42.5__0.1.2!backstage-plugin-mcp-actions-backend

2. You must see entries for any of the MCP tool plugins you installed as shown in the following code:

   ..... prior logs ....
   ======= Installing dynamic plugin oci://ghcr.io/redhat-developer/rhdh-plugin-export-overlays/red-hat-developer-hub-backstage-plugin-software-catalog-mcp-tool:bs_1.42.5__0.2.3!red-hat-developer-hub-backstage-plugin-software-catalog-mcp-tool
   	==> Copying image oci://ghcr.io/redhat-developer/rhdh-plugin-export-overlays/red-hat-developer-hub-backstage-plugin-software-catalog-mcp-tool:bs_1.42.5__0.2.3!red-hat-developer-hub-backstage-plugin-software-catalog-mcp-tool to local filesystem
   	==> Successfully installed dynamic plugin oci://ghcr.io/redhat-developer/rhdh-plugin-export-overlays/red-hat-developer-hub-backstage-plugin-software-catalog-mcp-tool:bs_1.42.5__0.2.3!red-hat-developer-hub-backstage-plugin-software-catalog-mcp-tool

Procedure

1. Consult your model documentation to confirm its support for tool calling.
2. If the current model does not support tool calling, change the model that your MCP client uses to a tool-calling compatible model.

Procedure

1. Verify the authentication token.

   1. Configure a static token for the RHDH MCP server.
   2. Set this token as the bearer token in your MCP client and make sure the authorization header is the configuration that specifies Bearer before the token, for example, Bearer <mcp_token>.

2. Check the MCP endpoint configuration.

   1. Confirm that the MCP server URL properly resolves, particularly if you are using a desktop client.
   2. Use legacy SSE endpoint if your MCP client requires it instead of the Streamable endpoint. (For more details, see the Configuration topic).

3. Verify your RHDH app-config.yaml file for duplicate backend entries and make sure that the indentation is accurate.

   1. Make sure that the configuration for the static token and MCP plugin sources go under an existing backend field, if one is present.
   2. If you are unsure, see the sample app configuration provided in the reference procedure.

1. Select an appropriate model for tool calling.

   1. Verify that the model has good support for tool calling.
   2. Make sure your model is not too small. We recommend a model with at least 7 billion parameters and a context window of 32k.

2. Refine your queries.

   1. Use more well-defined queries that limit the amount of data returned in the response from the tool.
3. If possible, increase the context window size of the model. We recommend at least 32k for these MCP tools.