Docs / AI assistant
AI assistant
Ask about what you're looking at, with the model you already use. See what's sent, and apply suggestions only after a dry run.
The assistant is a sidebar on the right: the spark button at the end of the tab bar, or Cmd/Ctrl+Shift+A. Drag its left edge to resize it. Nothing is sent until you ask.
Choose a provider
Turn it on in Settings → AI and choose what answers. It turns on by itself once the chosen provider is set up, for example when ANTHROPIC_API_KEY is set or a key is saved, unless you've switched it off.
A command-line tool you already have, needing no API key:
- Claude Code, with your own sign-in. It runs with every tool, MCP server and settings file turned off, in an empty temporary folder.
- Ollama.
- Any command that reads a prompt on standard input.
A provider with an API key:
- the Anthropic API with Claude, or an Anthropic-compatible gateway;
- any OpenAI-compatible endpoint, such as OpenAI, Ollama or LM Studio.
Claude in your own cloud account, with no gateway:
- Amazon Bedrock: requests are signed with your AWS credentials, from the environment or the AWS CLI profile you name (SSO included), or use a Bedrock API key.
- Vertex AI, or Claude models on Google Cloud, with an access token from your
gcloudsign-in.
Keys go to the system keychain, or come from ANTHROPIC_API_KEY, OPENAI_API_KEY or AWS_BEARER_TOKEN_BEDROCK. Cloud credentials stay in memory and are renewed before they expire.
Ask about what you're looking at
The page's context goes with your question:
- a resource, with its events, owners, Node and recent logs;
- the rows of a list;
- the Health problems;
- the topology.
Suggested questions fit the page, and moving to another page sends fresh context with the next question. Follow-up questions continue the conversation.
Ways in:
- a on a focused row, or Ask AI in any resource's actions menu or the topology panel;
- Ask AI on Health page rows, warning events and Timeline warnings;
- selected log lines in the log viewer;
- Ask AI what is wrong here in a Namespace's actions menu. It sends the unhealthy workloads and Pods, the last two hours of events, rollouts, restarts and scaling, and the two worst Pods in detail;
- Ask AI about all of these on the Health page, which treats every problem shown as one incident.
What's sent
Secret values and sensitive-looking environment values are left out, and every question shows exactly what was sent.
With an API provider, the assistant can also look things up while it answers, using read-only tools: list contexts, list or get resources, events, and logs if you allow them. Every lookup and what it returned is shown in the conversation. Settings → AI can turn lookups off.
From suggestion to change
The model only suggests; nothing in the cluster changes unless you do it. When an answer gives a change as a YAML manifest, Review and apply… under it opens New resource with that YAML and the conversation's context. A server-side dry run and diff come first, conflicts with other field managers can be reviewed and forced, and nothing changes until you apply.
To let coding agents on your computer use Klystiq's view of your clusters, see MCP server.