Skip to main content

Custom Reasoning Model

The Custom Reasoning Model integration lets a cluster use your own Anthropic API key instead of the default managed AWS Bedrock backend. Once connected, reasoning calls (investigation analysis, chat responses, recommendation generation) run against the Anthropic API using your key and your chosen Claude model.

This integration is Anthropic only. It is not a generic OpenAI-compatible endpoint: there is no endpoint-URL field, and provider values like Azure OpenAI or self-hosted OpenAI-compatible servers are not supported. For AWS Bedrock details see AWS Bedrock. For other model hosting, see Bring Your Own LLM.

When to Use​

  • You want reasoning billed to and running on your own Anthropic account.
  • You have compliance or procurement requirements that mandate a direct Anthropic relationship.
  • You want to pin a specific Claude model on a specific cluster.

If you don't have these constraints, the default AWS Bedrock backend is the simpler choice: there's no setup and it's tuned for the investigation workload.

Configuration​

In the OpsWorker portal:

  1. Go to Integrations → choose Custom Reasoning Model
  2. Provide:
    • Provider: select Anthropic (the only option)
    • API Key: your Anthropic API key (sk-ant-...)
    • Model: choose from the list of models, which is fetched dynamically from Anthropic using your key
  3. Save

On save, OpsWorker validates the key by calling Anthropic's models list. If the key is invalid (the call fails), the integration is not activated.

Scope​

Custom Reasoning Model integrations are stored per cluster. Different clusters can use different keys or models: for example, a regulated cluster can run on a dedicated Anthropic account while a development cluster uses the default Bedrock backend.

Considerations​

  • Quality varies by model. More capable Claude models produce noticeably better investigations than smaller, faster ones.
  • Latency affects time-to-investigation. A slow path lengthens investigations beyond the default sub-2-minute target.
  • Cost is on you. You pay Anthropic directly for the inference. OpsWorker does not mark up or bill for inference when you bring your own key.
  • No fallback by default. If your Anthropic key stops working, investigations on that cluster fail until it is restored; they do not silently fall back to the default backend.

Next Steps​