OpsPilot

Cast AI OpsPilot is an AI-powered assistant embedded in the Cast AI console. Ask operational questions about your Kubernetes clusters in plain English and get structured, actionable answers — without writing kubectl commands, navigating multiple dashboards, or running SQL queries.

Questions that previously required switching between dashboards and tools can be handled in a single interaction. OpsPilot understands your live cluster state, cost and savings data, workload events, audit logs, cluster health scores, and Cast AI documentation. It combines information from multiple data sources into a single response.

Access OpsPilot

OpsPilot is available to all Cast AI customers directly in the console — no separate configuration, installation, or onboarding required.

1. Log in to the Cast AI console.
2. Locate the OpsPilot chat interface in the console header. It is accessible from any page.
3. Type your question in the input field and press Enter.

OpsPilot authenticates you automatically using your existing Cast AI session.

How OpsPilot works

When you ask a question, OpsPilot analyzes your intent and routes the query to the right data sources. It can query multiple sources in parallel and combine the results into a single, coherent response.

OpsPilot covers three main data domains:

• Live cluster state — pods, nodes, services, deployments, statefulsets, daemonsets, HPAs, namespaces, workload events, and audit logs. OpsPilot queries your real-time cluster data and applies filtering, counting, and analysis.
• Cost and savings analytics — cluster costs, namespace-level cost breakdowns, workload-level costs, savings data, and optimization recommendations. Cost data updates at 5-minute intervals and can include trend charts.
• Cast AI documentation — how-to guides, feature explanations, and configuration references from the Cast AI docs. OpsPilot retrieves the most relevant content and synthesizes a targeted answer with source citations.

For multi-topic questions like "show me expensive pods with their resource requests," OpsPilot queries both the cost analytics and cluster state data simultaneously, then merges the results into a combined response.

What you can ask

OpsPilot covers five main capability areas. The table below summarizes what each covers and the kind of question it handles.

OpsPilot can also handle questions that span multiple areas simultaneously. "Which nodes are underutilized and what would I save by removing them?" draws on cluster state and savings data in a single response.

📘

Documentation index freshness

The documentation index updates periodically. Very recent documentation changes may take up to 24 hours to appear in OpsPilot answers.

Work with cluster context

OpsPilot determines which cluster your question is about automatically using these signals (in priority order):

1. Explicit mention — if you name a cluster in your question, OpsPilot uses that cluster.
2. "Current cluster" or "this cluster" — uses the cluster on the console page you're currently viewing.
3. Conversation context — if you asked about a cluster earlier in the session, subsequent questions default to it.
4. Page context — for the first question in a session, the cluster dashboard you're viewing is used.

To switch clusters mid-conversation, mention the new cluster by name. All subsequent questions will use the new cluster until you switch again.

If you're viewing one cluster's dashboard but your conversation is about a different cluster, OpsPilot displays a hint to prevent wrong-cluster mistakes. Minor typos in cluster names are corrected automatically with a confirmation note.

Follow-up questions

OpsPilot maintains conversation context within a session, so follow-up questions work naturally:

1. Ask an initial question: "Show me OOM kill events in the last 24 hours."
2. Follow up with: "Which workloads were affected most?"
3. Drill deeper: "What are the memory limits on those workloads?"

Interpret results

OpsPilot returns structured responses designed to help you take action, not just read data. Results land in different formats depending on what you asked:

• Tables give you scannable lists of resources, cost breakdowns, and comparisons — sorted and filtered so the most relevant rows surface first.
• Charts render cost trends and time-series data inline in the console, giving you the context to understand whether a cost spike is an anomaly or a pattern.
• Console links take you directly to the relevant page to act on what OpsPilot found — for example, to apply a recommendation or inspect a specific workload. Because OpsPilot is read-only, these links are how you get from answer to action.
• Recommendations surface optimization suggestions with estimated savings impact when your question is about cost or efficiency.
• Source citations accompany documentation-based answers with links to the relevant pages on docs.cast.ai.

Use cases

Incident investigation

When pods fail or latency increases, you typically need to check workload events, review recent changes in the audit log, and cross-reference with cluster state — a process that can take 15–20 minutes across multiple dashboards.

With OpsPilot, ask a single question like "What workload optimization events happened in the api-gateway namespace in the last 24 hours, and did any cause OOM kills?" and get the full picture: the change that was made, the events that followed, and the correlation between them.

Cost attribution

Engineering or finance teams often need cost breakdowns by namespace or workload to allocate cloud costs back to product teams. Instead of exporting data and building spreadsheets, ask OpsPilot "Give me a cost breakdown by namespace for the last 30 days across all clusters, ordered by total cost" and get a formatted result immediately. Follow up with "Which workload in the analytics namespace is most expensive?" to drill deeper.

Weekly reporting

Summarizing autoscaling activity or cost trends for stakeholders typically requires exporting event data, filtering, and building a summary. OpsPilot handles this in a single question: "How many autoscaling events happened last month and how many traffic surge events occurred?"

Onboarding new engineers

New team members can ask OpsPilot about Cast AI features and get targeted answers without leaving their workflow. Questions like "How does spot instance management work?" or "What policies control autoscaling?" return synthesized answers drawn from Cast AI documentation, with source links for further reading.

Data security and privacy

OpsPilot enforces strict data isolation and privacy:

• Organization-level isolation — every query is scoped to your organization. Data isolation is enforced at the database level, not just in the application layer. There is no path to access another organization's data.
• Input validation — all queries pass through multiple validation layers before processing, including pattern-based detection and machine-learning-based scanning for prompt injection attempts.
• Response sanitization — internal identifiers like cluster UUIDs are resolved to friendly names in responses. Internal system details are never surfaced.
• No model training on your data — your cluster data is never used to train AI models. OpsPilot queries your data in real time and returns results within your session.

Current limitations

Troubleshooting