What the AI does for you
- Investigate a problem - open the failing resource, press
⌘J, and ask why. The assistant reads its logs, events, manifest and Stack history. See AI Assistant. - Build or change a Stack - describe what you need and it proposes the add-ons, their order and starting values. See Building a Stack.
- Explain a firing alert - an alert firing gets an AI analysis with the root cause and recommended actions. See Alert-triggered analysis.
- Find problems before they page you - scheduled scans raise findings with a root cause and a fix. See AI Insights.
- Review pull requests - a verdict and findings on every pull request, and answers to
@ankraaimentions. See AI code review. - Work tickets - agents pick up tickets, draft a plan, and execute it once a person approves. See AI Board, AI Agents and AI Automations.
What it can change without asking
Five controls bound what the AI may do, each with a safe default. A chat write that runs without a confirmation card is recorded in the audit log with the autonomy setting that allowed it.
Whatever Auto or Autonomous allows, these always wait for a person:
- Cluster lifecycle - creating, scaling, upgrading, stopping, moving or deleting a cluster, its node pools or its control plane.
- Removals - a Stack change that would remove or replace a deployed add-on or manifest, or delete a Stack.
- High-risk and irreversible writes, and a Stack change whose validation failed.
- Alert silences - muting an alert always needs a person.
kube-*, ankra-*, and system namespaces such as cert-manager and vault) are refused outright in every mode, and Kubernetes Secret values are withheld from the model. The exception to all of this is a remediation per-action rule set to Run automatically: that action runs with no approval card, so set it only for an action you would approve every time.
Stopping the AI
Two organisation-wide stops, from least to most disruptive. Both take effect immediately and need organisation admin.- Pause autonomous actions - chat, diagnosis and read-only investigation keep working, but nothing changes without a person approving it. It switches off the remediation policy and pauses every running agent; resuming restores exactly what it switched off.
- Stop all AI - the emergency stop. It cancels every running AI session and agent run, expires pending approvals, and blocks chat for everyone until an admin resumes it.
ankra ai autonomy resume and ankra ai autonomy start-all undo them. See AI Autonomy.
Where you can use it
ankra chat accepts --mode ask or --mode agent; without it, the chat runs in Agent mode. Every surface uses the same tools and the same permission checks, and answers with the permissions of the person asking.
Models and cost
On Ankra’s default models, chat routes each question to one of three tiers: Expert (Kimi K3), Think (GLM 5.3) and Quick (Claude Haiku 4.5). Usage is priced per request and counted against a free monthly allowance per organisation - USD 10 on Free, USD 50 on Startup, USD 200 on Enterprise - and an optional daily spend cap you set yourself under Billing → Overview. When the allowance runs out, platform-funded AI pauses until the next month, and the rest of Ankra keeps working. With your own OpenRouter, Anthropic or OpenAI-compatible key, usage bills to your account and never draws on the allowance. See Models and Cost and Bring your own OpenRouter API key.How work flows
Everything the AI does in the AI section converges on the Board. Work arrives from four directions: a firing alert spawns remediation, a proactive insight is filed, an automation reaches an approval gate, or a person files a ticket by hand.Get started
Press⌘J on a cluster page and ask what needs attention. To set up the rest of the AI section, an admin opens Guided setup from AI → Home.