AI, with the boundaries in view
An investigation assistant.
You stay in control.
Ask a question about your cluster. Inspect the tool calls. Review the suggested fix before you act.
What happens when you ask?
- The assistant starts with context. A recent cluster snapshot provides workload status, events and capacity information.
- It gathers evidence. Read-only tools inspect the resources and APIs your installation permits. Missing permissions and incomplete coverage limit the answer.
- It proposes a next step. You receive an explanation and, where useful, commands or manifests to review. The agent has no path to apply cluster changes.
Where your data goes
The selected provider receives the conversation, cluster context and tool results needed for the run. That data can contain sensitive workload information.
Requests go to the Ollama server you configure. Operate it within your infrastructure if you want inference to remain there. Model downloads, compute and network setup are your responsibility.
Requests go directly to the provider endpoint using your credentials. They are governed by your provider agreement and settings. There is no Berth-hosted inference proxy.
What the tools can read
Generic resource tools block Secret and ConfigMap contents and expose limited metadata and replica counters for other objects. Raw logs and descriptions are disabled by default; an operator must explicitly enable them.
Credential-pattern filtering reduces some disclosures, but it cannot guarantee that arbitrary prompts, events or diagnostic text contain no secrets. Review data egress before enabling a model or raw diagnostics.
What remains your responsibility
- Check the diagnosis. A model can misunderstand evidence or be misled by text in a workload. Visible tool calls help you inspect its work; they do not prove an answer is correct.
- Review changes. Suggested commands can be wrong or disruptive. Use your normal review and deployment process.
- Keep historical monitoring. A live snapshot does not establish long-term behavior or prove that a node is safe to remove.
- Restrict access. The dashboard is for trusted cluster administrators. All authenticated users share its Kubernetes permissions.
Model costs and limits
Berth adds no markup to provider usage. Ollama uses your own compute; cloud providers bill your account. Enterprise is required to use cloud models, along with authentication and persistent usage storage.
Token budgets use estimated reservations before a run. Actual multi-round usage can exceed the reservation, so they are not a guaranteed financial ceiling. Use provider-side spending controls and reconcile the usage ledger against your provider’s bill.
Ready to try it?
Start with the dashboard, then connect a model when you need help investigating.