API keys and model defaults for every AI feature in the CMS — chat, agents, field generation, SEO optimisation, translation, interactives.
Where it is
Settings → AI (/admin/settings?tab=ai).
This tab has two sections: AI Providers (the API keys) and AI Model Defaults (which model goes to which feature).
AI Providers
Parse the feature via an API key from a provider. Currently supported:
| Provider | What for |
|---|---|
| Anthropic | Claude — used by default for everything (chat, agents, field generation, SEO, GEO, translate) |
| OpenAI | GPT — alternative path, used if you prefer non-Anthropic routing |
Keys are stored per site, not per account. Two sites in the same org can use different keys if you want isolated billing. Keys never leave the server — editors with editor or viewer roles cannot read them back.
Environment fallback
If a key isn't set per-site, the CMS falls back to ANTHROPIC_API_KEY / OPENAI_API_KEY from the environment. Useful in dev when you don't want to paste keys into every test site, but don't rely on this in production — a missing per-site key hides a forgotten configuration.
AI Model Defaults
Choose the model for each feature category. Different tasks want different models:
| Slot | Recommended | When to deviate |
|---|---|---|
| Content model | Haiku or Sonnet | Use Opus for premium sites where field generation quality matters more than cost |
| Chat model | Sonnet | Opus for complex tool-calling; never Haiku (tool-calling degrades) |
| Premium model | Opus | Used for Brand Voice interviews and long-form generation where one-shot quality is critical |
| Code model | Sonnet | Used for generate_interactive. Haiku produces unreliable JavaScript |
Also configurable per slot:
- Max tokens — maximum output per request (typical: 4096 for fields, 16384 for chat, 8192 for code)
- Temperature — creativity knob. 0.0 for factual (SEO optimiser), 0.7 for creative (content generation), 1.0 for brainstorming (never for production content)
- Max iterations — how many tool calls the chat can chain per turn. Default 25. Raise for complex multi-step research flows.
Per-feature overrides
Some features can override the model at the feature level:
- Agents — each agent config can pin a model override
- Prompts — the AI Prompts tab can override model per prompt template
- Workflows — each workflow node can pin its own model
Overrides let you use Opus for one high-stakes agent while keeping every other feature on Sonnet for cost control.
Rate limits & budgets
This tab doesn't enforce budgets directly — that's handled by the agent budget setting (per agent) and cockpit (global caps). See Cockpit for site-wide AI cost control.
Related
- AI agents — configure individual agents
- Cockpit — global temperature, token budget, cost ceilings
- AI Prompts — prompt template overrides
- Brand Voice — the voice injected into every AI prompt