/next/organisation/ai-models). The page has these sections: AI provider, Members’ own AI keys, Chat models, Custom endpoints, When a custom endpoint fails during a chat turn, Models for each AI feature, How Ankra AI replies and System prompt version. AI usage and billing, at the top, opens your AI spend on Plan and billing.
Provider modes
The AI provider section lists four credentials, one row each.
A row you have not set up reads Not set up; a saved key shows only its last characters (Key ending …).
Members can also add API keys of their own to their profile, for the AI work they start themselves, while Members can use their own AI keys is on under Members’ own AI keys. See Your Own AI Keys.
Credentials are stored independently, and more than one can be saved at once. A saved credential starts serving the calls only it can serve as soon as it is saved. Prefer this credential, in a row’s options menu, decides which key runs where two of them could serve the same call, such as an Anthropic key and an OpenRouter key; that row is then marked Preferred. The AI provider section states which credential runs what, and who pays for it.
Default models
Every organisation has the same three tiers. The chat model picker selects between them, and Auto routes each question to the right one - short questions to Quick, everyday work to the default, and architecture, security, or image questions to Expert. Which model runs each tier depends on the keys your organisation has saved under AI provider:
Every model in the table has a 1M-token context window except Claude Haiku 4.5, which has 200k; Chat models shows each model’s window. An organisation with both an OpenRouter key and an Anthropic key gets the OpenRouter column, and its Anthropic key runs Quick.
The table is the default catalog, which an organisation gets until it edits Chat models. A catalog your organisation has edited keeps the models it was saved with, whatever keys you add or remove, until someone chooses Go back to Ankra’s defaults….
Auto sends any question with an attachment to Expert. On your own OpenRouter key, Think cannot take one: GLM 5.3 does not accept images.
The DevOps AI model by Ankra
Beside the three tiers, every catalog carries DevOps AI: Ankra’s own model, served on Ankra’s own servers and preloaded with the Ankra CLI skills, so it answers in Ankra’s terms - theankra CLI, stacks, GitOps, applications on Ankra Pipelines - without being told about them. It is text only, keeps a 128k-token context, and calls tools like the other models. Pick it in the chat model picker, or set it as a function’s model; Auto never routes to it on its own.
It comes with an allowance of its own. Every plan, the free tier included, has $50 a month on the DevOps AI model, separate from the free monthly allowance. Past that it bills per token, like any model Ankra serves. The DevOps AI model is also what answers once your organisation runs out: when the free monthly allowance is spent, or the organisation’s credit is gone, AI features switch to the DevOps AI model while its own allowance has room instead of stopping, and the conversation shows a notice when they do. Add your own model key to run on your own account instead.
These are defaults, not limits. Under Chat models you can edit any tier, point it at a different model, add your own, or reset the catalog back to the Ankra defaults.
Models per function
AI work that runs outside chat has its own model setting per function on Ankra AI → Models and provider, in the section Models for each AI feature. A function with nothing selected follows its default tier, so it moves with your catalog: point a tier at a different model and every function on that tier follows it.
Two kinds of AI work are not in this list and do not follow your provider: the analysis Ankra writes when an alert fires, and the scheduled health check. Both run on Ankra’s own access, Groq first. Which model reads it says what they are sent.
An organisation that has changed nothing therefore runs every function in this list on Claude Haiku 5.5 on Ankra’s managed access. On its own OpenRouter key the Expert functions run on Kimi K3 and the Think functions on GLM 5.3; on its own Anthropic key, on Claude Opus 5.5 and Claude Sonnet 5. See Default models.
In the portal this list is the Models for each AI feature section of Models and provider: one row per feature, each with a Model for feature picker. Pick any model in your catalog for a feature, or put it back on its default to follow the matching tier in Chat models. Only owners and admins can change a feature’s model. From the terminal,
ankra ai lanes list shows what each function runs on, and ankra ai lanes set <lane> <model> changes it - it also accepts an exact OpenRouter model id.
Change the provider
ankra ai models edits the model catalog, ankra ai lanes the models per function, and ankra ai endpoints the custom endpoints. The sections below describe the same steps in the portal, on Ankra AI → Models and provider.
Configure Anthropic
1
Save the key
In AI provider, press Set up… on Your Anthropic key (Replace key… once one is saved). The dialog Save your Anthropic API key asks for the key you made in the Claude Console at platform.claude.com. Paste it into Anthropic API key and press Save key. Ankra sends it a one-token check before it saves it.The row now shows Key ending and the key’s last characters.
2
Prefer it where keys overlap
If you also hold an OpenRouter key, open the Anthropic row’s options menu and press Prefer this credential to decide which of the two runs the calls both could serve. The row is now marked Preferred. The menu offers it only once a key is saved.
Configure OpenRouter
With your organisation’s own OpenRouter API key active, Ankra serves the OpenRouter default models (Kimi K3 and GLM 5.3, where Ankra’s managed access runs Claude Haiku 5.5) and the automatic failover chain through your OpenRouter account - including background and autonomous AI sessions - and your usage never draws on the free monthly allowance. The key is checked against OpenRouter before it is saved, and stored only in the secret store. Follow the step-by-step guide: Bring Your Own OpenRouter API Key.Configure an OpenAI-compatible endpoint
1
Save the configuration
In AI provider, press Set up… on Your OpenAI-compatible endpoint (Change… once one is saved). In Set up an OpenAI-compatible endpoint, fill in:
- Base URL - the address that ends at the OpenAI
/v1root, for examplehttps://my-litellm.internal/v1. Anhttp://address shows This address is not encrypted. - Model - the exact model id the endpoint serves
- API key - the endpoint’s key. Ankra sends it as a bearer token, stores it encrypted and never shows it again.
2
Point chat models at it
In Custom endpoints, press Add endpoint, give it a Name, Base URL and API key, and save. Then in Chat models, press Add model, choose the endpoint under Runs on, and pick from Models the endpoint lists or type the Model id. An endpoint runs only the models you add to it.
The endpoint serves the chat models you list under it, plus Stack summaries. The built-in Expert and Think tiers stay on their own models unless you point them at your endpoint under Chat models, and AI Insights, CI/CD pipeline generation and deploy analysis fall back to the Ankra default for tool-using steps. Each model in Chat models shows which credential runs it.
When a custom endpoint fails during a chat turn
This section decides what happens when your own endpoint fails in the middle of a chat turn:- Carry on with Ankra’s managed access - the turn continues on Ankra’s managed access, and the chat says that it did.
- End the turn with an error - the conversation never leaves your own endpoint once a turn has started on it.
How Ankra AI replies
Reply style sets how Ankra AI words and sizes its replies in chat, sessions and the work it runs for you: Balanced (results with brief context, and summaries where they help), Short (replies that lead with the result and leave out preamble), Explained (the reasoning and the platform context behind each result) or Teaching (the Kubernetes and platform ideas behind each result, briefly, as it goes). It never changes what Ankra AI is allowed to do. System prompt version shows the instructions Ankra AI starts from. Ankra manages them; Ankra support can pin your organisation to one version. Contact support to change it.Usage and the free allowance
Ankra meters AI requests - chats, agent sessions, AI Insights, CI/CD file generation, code reviews - per organisation, priced in USD from the tokens each model used. The analysis Ankra writes when an alert fires is not metered to your organisation. Settings → Plan and billing (/next/organisation/billing) shows month-to-date AI usage and how much of the free monthly allowance remains; AI usage and billing on Models and provider opens it there.
The allowance covers usage on Ankra’s managed model access: the default when your organisation has no provider of its own, and anything your own key cannot serve. Usage on your own OpenRouter, Anthropic or OpenAI-compatible key never draws on it.
Defaults can change; Settings → Plan and billing always shows your organisation’s live allowance. For more headroom, contact support.
When the allowance runs out
AI features that would spend Ankra’s key pause for the rest of the calendar month (UTC). Those requests fail with the stable error codeAI_ALLOWANCE_EXHAUSTED, which you can match on in automation. The rest of the platform, and AI on your own key, keep working. To resume before the reset, bring your own OpenRouter API key or another provider.
When your own key is out of credits or rejected
When your organisation’s own OpenRouter or Anthropic account refuses a turn, the turn ends with a message naming the provider and the fix, and a stable error code you can match on in automation. No retry clears it; whoever manages your organisation’s AI settings does.
For a custom OpenAI-compatible endpoint, the message names the provider behind your organisation’s custom endpoint instead, and the fix is on the endpoint.
Daily spend cap vs monthly allowance
The daily AI spend cap is a separate brake you set yourself on Settings → Plan and billing:Permissions
Reading Models and provider needs a role that may use Ankra AI; without one the page reads Your role cannot see Ankra AI’s settings. Saving credentials, preferring one, and changing models, the reply style or a feature’s model need a role that manages Ankra AI: owners and admins. Everyone else sees the settings read-only, and the page says Your role can look but not change. People who manage Ankra AI, owners and admins, change these settings.Other AI settings
Besides Models and provider, the Configuration rows of the Ankra AI menu are:- Skills and tools - skills and custom tools; see AI Skills and Custom Tools
- Autonomy and approvals - the stops, auto-remediation and Proactive AI Insights; see AI Autonomy
- AI workspaces - see AI Workspaces
- Memory - see AI Memory and Feedback
- Feedback - response ratings; see AI Memory and Feedback
- Gateway connections - Slack, Microsoft Teams and code hosts; see AI Connections