Guardrails

A guardrails object is a reusable rule set owned by one workspace. Attach it to any number of keys; editing it changes all of them at once.

Fields

Name
Allowed models: empty means every enabled model. Otherwise the key may call only these; anything else gets 403 model_not_allowed. Aliases (libra/smart, libra/fast, libra/cheap) count as models: listing an alias permits whichever member of its chain answers. For a model you name directly, its fallback is used only if that fallback is on the list too; if nothing on the list can answer, the request gets 400 no_eligible_provider.
Blocked providers: removed from every fallback chain for this key. If a chain becomes empty the call returns 400 no_eligible_provider. Use this to keep a key on ASIX-hosted models only.
Max tokens per request: LIBRA clamps the client's max_tokens down to this; the request is not rejected.
Store request bodies: off by default. When on, prompts and completions are kept 7 days and visible in Logs to people who can see the key. Turn on only for debugging.

Owner, Superadmin and the workspace's Admin. Members can read.

Every workspace can name one guardrails object as default (Settings > Workspace). New keys in that workspace get it automatically.

Only with zero attached keys. Detach first (key detail > Guardrails > None).

Production key: allowed models = the one or two you actually use; blocked providers = anything outside Indonesia if data residency matters; max tokens = your real ceiling; logging off.
Debug key: same models, logging on, short expiration.