Skip to content

Remote MCP Servers

Warning

Remote MCP servers are in Tech Preview stage.

Configuration Change

The MCP server configuration format has been updated. The url field must now be specified inside the config section instead of at the top level. The old format (with url at the top level) is still supported for backward compatibility but will log a migration warning. Please update your configurations to use the new format.

SSE Mode Deprecation

The SSE (Server-Sent Events) transport mode for MCP is being deprecated across the MCP ecosystem. We strongly recommend using streamable-http mode for new MCP server integrations. SSE mode support is maintained for backward compatibility but may be removed in future versions.

HolmesGPT can integrate with remote MCP servers using either streamable-http (recommended) or SSE (deprecated) transport modes. This capability enables HolmesGPT to access external data sources and tools in real time. This guide provides step-by-step instructions for configuring HolmesGPT to connect with remote MCP servers.

Transport Modes

HolmesGPT supports two MCP transport modes:

  1. streamable-http (Recommended): Modern transport mode that uses HTTP POST requests with JSON responses. This is the preferred mode for new integrations.
  2. sse (Deprecated): Legacy transport mode using Server-Sent Events. Maintained for backward compatibility only.

Configuring Transport Mode

The transport mode and URL are specified in the config section of your MCP server configuration:

mcp_servers:
  my_server:
    description: "My MCP server"
    config:
      url: "http://example.com:8000/mcp/messages"  # Path depends on your server (could be /mcp, /mcp/messages, etc.)
      mode: streamable-http  # Explicitly set the mode
      headers:
        Authorization: "Bearer token123"

URL Path Variability

Different MCP servers may use different endpoint paths. Common examples include: - /mcp/messages - Used by some servers - /mcp - Used by other servers (e.g., Azure Kubernetes MCP) - Custom paths as defined by your server

The streamable-http client automatically handles the protocol regardless of the path. Consult your MCP server's documentation to determine the correct endpoint URL.

If no mode is specified, the system defaults to sse for backward compatibility. However, this default will be deprecated in the future, and you should explicitly set mode: streamable-http or mode: sse for new and old servers.

URL Format

  • Streamable-HTTP: URL should point to the MCP server endpoint. The exact path depends on your server configuration:
  • Some servers use /mcp/messages (e.g., http://example.com:8000/mcp/messages)
  • Others use /mcp (e.g., http://example.com:3333/mcp)
  • The streamable-http client automatically handles POST requests and responses at the provided URL
  • Check your MCP server documentation for the correct endpoint path
  • SSE: URL should end with /sse (e.g., http://example.com:8000/sse). If the URL doesn't end with /sse, HolmesGPT will automatically append it.

Example: MCP server configuration

When using the standalone Holmes Helm Chart, update your values.yaml:

mcp_servers:
  mcp_server_1:
    # human-readable description of the mcp server (this is not seen by the AI model - its just for users)
    description: "Remote mcp server using streamable-http"
    config:
      url: "http://example.com:8000/mcp/messages"  # Path may vary: /mcp, /mcp/messages, or custom path
      mode: streamable-http  # Explicitly set the preferred mode
      headers:
        Authorization: "Bearer {{ env.my_mcp_server_key }}"  # You can use holmes environment variables as headers
    llm_instructions: "This server provides general data access capabilities. Use it when you need to retrieve external information or perform remote operations that aren't covered by other toolsets."

Apply the configuration:

helm upgrade holmes holmes/holmes --values=values.yaml

When using the Robusta Helm Chart (which includes HolmesGPT), update your generated_values.yaml (note: add the holmes: prefix):

enableHolmesGPT: true
holmes:
  mcp_servers:
    mcp_server_1:
      # human-readable description of the mcp server (this is not seen by the AI model - its just for users)
      description: "Remote mcp server using streamable-http"
      config:
        url: "http://example.com:8000/mcp/messages"  # Path may vary: /mcp, /mcp/messages, or custom path
        mode: streamable-http  # Explicitly set the preferred mode
        headers:
          Authorization: "Bearer {{ env.my_mcp_server_key }}"  # You can use holmes environment variables as headers
      llm_instructions: "This server provides general data access capabilities. Use it when you need to retrieve external information or perform remote operations that aren't covered by other toolsets."

Apply the configuration:

helm upgrade robusta robusta/robusta --values=generated_values.yaml --set clusterName=<YOUR_CLUSTER_NAME>

SSE (Deprecated - Use Only for Legacy Servers)

When using the standalone Holmes Helm Chart, update your values.yaml:

mcp_servers:
  mcp_server_legacy:
    description: "Legacy MCP server using SSE (deprecated)"
    config:
      url: "http://example.com:8000/sse"  # Must end with /sse
      mode: sse  # Explicitly set, though this is deprecated
    llm_instructions: "Legacy server using deprecated SSE transport."

  mcp_server_2:
    description: "MCP server that runs in my cluster"
    config:
      url: "http://<service-name>.<namespace>.svc.cluster.local:<service-port>/sse"  # SSE endpoint must end with /sse
      mode: sse  # Explicitly set SSE mode (deprecated)
      headers:
        key: "{{ env.my_mcp_server_key }}"  # You can use holmes environment variables as headers for the MCP server requests.
    llm_instructions: "This is a cluster-local MCP server that provides internal cluster data and operations. Use it for accessing cluster-specific information, internal services, or custom tooling deployed within the Kubernetes environment."

Apply the configuration:

helm upgrade holmes holmes/holmes --values=values.yaml

When using the Robusta Helm Chart (which includes HolmesGPT), update your generated_values.yaml (note: add the holmes: prefix):

enableHolmesGPT: true
holmes:
  mcp_servers:
    mcp_server_legacy:
      description: "Legacy MCP server using SSE (deprecated)"
      config:
        url: "http://example.com:8000/sse"  # Must end with /sse
        mode: sse  # Explicitly set, though this is deprecated
      llm_instructions: "Legacy server using deprecated SSE transport."

    mcp_server_2:
      description: "MCP server that runs in my cluster"
      config:
        url: "http://<service-name>.<namespace>.svc.cluster.local:<service-port>/sse"  # SSE endpoint must end with /sse
        mode: sse  # Explicitly set SSE mode (deprecated)
        headers:
          key: "{{ env.my_mcp_server_key }}"  # You can use holmes environment variables as headers for the MCP server requests.
      llm_instructions: "This is a cluster-local MCP server that provides internal cluster data and operations. Use it for accessing cluster-specific information, internal services, or custom tooling deployed within the Kubernetes environment."

Apply the configuration:

helm upgrade robusta robusta/robusta --values=generated_values.yaml --set clusterName=<YOUR_CLUSTER_NAME>

Example: Working with Stdio MCP servers

MCP currently supports three transport mechanisms: stdio, Server-Sent Events (SSE - deprecated), and Streamable HTTP. HolmesGPT supports both streamable-http (recommended) and SSE (deprecated) transport modes. However, many existing MCP servers—such as Dynatrace MCP—rely exclusively on the stdio transport. To overcome this incompatibility, tools like Supergateway can act as a bridge by converting stdio-based MCPs into streamable-http or SSE-compatible endpoints.

Prefer Streamable-HTTP

When using Supergateway or similar tools, configure them to use streamable-http mode instead of SSE for better compatibility and future-proofing.

For this demo we will use: - Dynatrace MCP - Supergateway - runs MCP stdio-based servers over SSE

Check out supergatway docs to find out other useful flags.

See it in action

1. Run stdio MCP as SSE

This command runs the Dynatrace MCP server locally via Docker using Supergateway to wrap it with SSE support. Credentials (e.g., API keys) should be stored in a .env file passed to Docker using --env-file. you can change "npx -y @dynatrace-oss/dynatrace-mcp-server@latest /" to your specific MCP.

docker run --env-file .env -it --rm -p  8003:8003 supercorp/supergateway \
--stdio "npx -y @dynatrace-oss/dynatrace-mcp-server@latest /" \
--port 8003 \
--logLevel debug

Once the container starts, you should see logs similar to:

[supergateway] Starting...
[supergateway] Supergateway is supported by Supermachine (hosted MCPs) - https://supermachine.ai
[supergateway]   - outputTransport: sse
[supergateway]   - Headers: (none)
[supergateway]   - port: 8003
[supergateway]   - stdio: npx -y @dynatrace-oss/dynatrace-mcp-server@latest /
[supergateway]   - ssePath: /sse
[supergateway]   - messagePath: /message
[supergateway]   - CORS: disabled
[supergateway]   - Health endpoints: (none)
[supergateway] Listening on port 8003
[supergateway] SSE endpoint: http://localhost:8003/sse
[supergateway] POST messages: http://localhost:8003/message

This will run dynatrace MCP server as a pod in your cluster. credentials are passed as env vars.

apiVersion: v1
kind: Pod
metadata:
  name: dynatrace-mcp
  labels:
    app: dynatrace-mcp
spec:
  containers:
    - name: supergateway
      image: supercorp/supergateway
      env:
        - name: DT_ENVIRONMENT
          value: https://abcd1234.apps.dynatrace.com
        - name: OAUTH_CLIENT_ID
          value: dt0s02.SAMPLE
        - name: OAUTH_CLIENT_SECRET
          valueFrom:
            secretKeyRef:
              name: dynatrace-credentials
              key: client_secret
      ports:
        - containerPort: 8003
      args:
        - "--stdio"
        - "npx -y @dynatrace-oss/dynatrace-mcp-server@latest /"
        - "--port"
        - "8003"
        - "--logLevel"
        - "debug"
      stdin: true
      tty: true
---
apiVersion: v1
kind: Service
metadata:
  name: dynatrace-mcp
spec:
  selector:
    app: dynatrace-mcp
  ports:
    - protocol: TCP
      port: 8003
      targetPort: 8003
  type: ClusterIP

2. Add MCP server to holmes config

With the MCP server running in SSE mode, we need to let HolmesGPT know of the mcp server. Use this config according to your use case.

Configuration:

Use a config file, and pass it when running cli commands.

custom_toolset.yaml:

mcp_servers:
  mcp_server_1:
    description: "Dynatrace observability platform. Bring real-time observability data directly into your development workflow."
    config:
      url: "http://localhost:8003/sse"
    llm_instructions: "Use Dynatrace to analyze application performance, infrastructure monitoring, and real-time observability data. Query metrics, traces, and logs to identify performance bottlenecks, errors, and system health issues in your applications and infrastructure."

You can now use Holmes via the CLI with your configured MCP server. For example:

holmes ask -t custom_toolset.yaml  "Using dynatrace what issues do I have in my cluster?"

Alternatively, you can add the mcp_servers configurations to ** ~/.holmes/config.yaml**, and run:

holmes ask "Using dynatrace what issues do I have in my cluster?"

When using the standalone Holmes Helm Chart, update your values.yaml:

mcp_servers:
  mcp_server_1:
    description: "Dynatrace observability platform. Bring real-time observability data directly into your development workflow."
    config:
      url: "http://dynatrace-mcp.default.svc.cluster.local:8003"
    llm_instructions: "Use Dynatrace to analyze application performance, infrastructure monitoring, and real-time observability data. Query metrics, traces, and logs to identify performance bottlenecks, errors, and system health issues in your applications and infrastructure."

Apply the configuration:

helm upgrade holmes holmes/holmes --values=values.yaml

When using the Robusta Helm Chart (which includes HolmesGPT), update your generated_values.yaml (note: add the holmes: prefix):

enableHolmesGPT: true
holmes:
  mcp_servers:
    mcp_server_1:
      description: "Dynatrace observability platform. Bring real-time observability data directly into your development workflow."
      config:
        url: "http://dynatrace-mcp.default.svc.cluster.local:8003"
      llm_instructions: "Use Dynatrace to analyze application performance, infrastructure monitoring, and real-time observability data. Query metrics, traces, and logs to identify performance bottlenecks, errors, and system health issues in your applications and infrastructure."

Apply the configuration:

helm upgrade robusta robusta/robusta --values=generated_values.yaml --set clusterName=<YOUR_CLUSTER_NAME>

After the deployment is complete, you can use HolmesGPT and ask questions like Using dynatrace what issues do I have in my cluster?.