Skip to content
Docs · Documentation

Kubyl Mobile

AI agents in the app

Chat with an AI agent about a cluster from Kubyl Mobile: consent, entry points, what it can read and what stays on the server.

The app is a thin chat client. The agents (Claude, Codex, Gemini CLI, GitHub Copilot, goose, OpenCode or a custom one) run on your Kubyl Server and read the cluster through read-only tools with your own rights. They cannot change anything. The server owner installs and signs agents in, never a device; see AI agents on the server.

When it appears

Only when the server reports the agents capability and at least one agent is ready (installed and signed in). Otherwise the app shows nothing and asks nothing. An agent the owner has not signed in is listed in the picker with “The server owner must sign this agent in.”

Before your first thread with an agent a sheet shows the server's disclosure and the agent's provider (who receives the data: logs, object YAML with Secrets masked, events, metrics, alerts and what you type). Agree or Not now. Consent is stored per server and agent, and asked again if the provider or the disclosure text changes. Revoke it under Settings, AI agents (the section appears once you have agreed); the next message to that agent asks again.

Starting a thread

  • The cluster bar has the thread list and New thread; More, Agents, Ask agent (“Chat about this cluster”) does the same.
  • Ask from context: an alert page, any object's page (the icon in the top bar, so it also works on read-only clusters), a Flux object and the pod logs. The chat starts with a removable chip naming that object, alert or log (with the container) so the agent knows what you mean.
  • With several agents ready you pick one first. The thread list shows this cluster's threads first, and a menu closes a thread.

The chat

  • The transcript shows your messages, the agent's answers (Markdown: headings, lists, code blocks with copy, quotes, bold, italic, inline code and links; tables and HTML show as plain text, and only http(s) and mailto links are live), Kubyl tool cards with collapsible output (you see the first part of a long output), collapsed thoughts, a plan bar and notices.
  • The agent may ask you something: a choice, a short form or whether to open a link. Requests to run its own tools, a terminal or file access are refused automatically; the thread says “this server only allows Kubyl's cluster tools”.
  • Send and Stop. Banners explain a thread that is starting, needs a sign-in (ask the server owner), failed or ended; Start a new thread brings your last prompt back. Prompts longer than the server's limit are refused before sending (default 8 KiB).
  • A thread belongs to your device. Threads are not saved: a server restart ends every conversation. See the server's limits (threads per device, running turns, idle time) in AI agents on the server.

Something missing or wrong? Open an issue on GitHub.