Skip to main content

AI Operations Workspace

The AI workspace is the primary natural-language operating surface in KubeVision. It combines the current page context with live Kubernetes tools so answers can be grounded in the selected cluster instead of relying only on general model knowledge. It is available as a side panel or a full-tab chat workspace.

What It Can Do​

  • Read one Kubernetes resource or list resources by kind and label
  • Read bounded recent Pod logs, including a previous container instance
  • Summarize cluster health and identify Pods outside Running/Succeeded states
  • Query a configured Prometheus instance
  • Prepare create, update, patch, and delete operations for approval
  • Stream explanations and tool results while a multi-step investigation runs

Configure a Provider​

Open Settings > AI Assistant. Administrators can enter an endpoint and credentials, fetch available models, and test the configuration. Credentials are encrypted at rest and are not returned by the API after saving.

KubeVision uses OpenAI-compatible chat and model APIs. This supports providers such as OpenAI, OpenRouter, DeepSeek, Qwen, and compatible self-hosted gateways. Compatibility depends on the provider implementing streaming and tool calling correctly.

Only configure a service you trust. User prompts, page context, resource data, log excerpts, metrics results, and tool outputs used in a conversation are sent to that provider. A self-hosted compatible endpoint can be used when cluster data must remain on your network.

Guarded Tool Actions​

Read operations are restricted to resources the current user can access. Mutating tools use the same RBAC rules as the regular UI and pause for explicit confirmation before execution. The server re-checks permission when the user confirms, so an expired approval or changed role cannot reuse the earlier authorization decision. Approved mutations generate audit entries.

Review the target cluster, namespace, resource, manifest or patch, and generated arguments before confirming. Model output is not a substitute for a rollout, backup, disruption, or data-retention plan.

When a request requires several Kubernetes steps, the assistant may continue after tool results are returned. Keep the chat open until it reports completion or a concrete failure; interrupted confirmation or expired authentication can stop the sequence.

Display Modes​

Use the layout control to switch between the side panel and full-tab workspace. The conversation remains available when switching modes. Full-tab mode is better for long investigations; the panel keeps the current resource visible.

Conversations and drafts are stored in the browser, scoped to the signed-in user, and restored after logout or a later login on the same browser. They are not synchronized through the KubeVision server or shared with another browser. Clear the site's browser data to remove them. Avoid leaving sensitive cluster content on a shared device.

Current Boundaries​

  • The assistant does not run autonomous background remediation.
  • It operates on one selected cluster per conversation turn.
  • Only the tools listed above are available; terminal execution and arbitrary network access are not exposed to the model.
  • A conversation run has bounded tool iterations and bounded log/tool output to prevent runaway loops and oversized model context.