Skip to main content
Agents run against models from external providers (OpenAI, Anthropic, Google) plus Nora-managed model tiers. Each provider that agents call needs an API key at the workspace level.

​
Where keys live

  • Workspace-scoped — keys are per workspace, not per user.
  • Encrypted at rest — never returned via UI or API. The app can’t read your key back.
  • Per-workspace usage attribution — keeping keys per workspace makes cost attribution clean on the provider dashboard.

​
Adding a key

Settings → Models → Providers → Add key:
  1. Pick the provider (OpenAI, Anthropic, Google, …).
  2. Paste the key. Input is masked.
  3. Save.
Or via CLI:
See nora providers for CLI details.

​
Which providers Nora supports

  • OpenAI — GPT models.
  • Anthropic — Claude models.
  • Google — Gemini models.
  • Nora managed — Nora-hosted models (no external key required; billed by Nora).
More providers ship regularly. The providers list in the app shows the current set.

​
Rotating a key

Re-add the key with a new value. Old key is replaced atomically:
  • New runs use the new key immediately.
  • In-flight runs finish on the old key (up to a minute for the network to settle).
Verify by watching for auth failures over the next few minutes:

​
Deleting a key

Removes the key. Any Agent using that provider will fail on next run until you set a new key.

​
Provider readiness check

Before publishing a Flow, verify all its models have provider keys:
Prints a table:
CI can gate publishes on this check.

​
Provider vs. Nora-managed models

  • Provider keys — you pay the provider directly, on their meter.
  • Nora-managed — Nora hosts the model, bills you on Nora’s invoice. Simpler billing; sometimes cheaper.
Choose per Agent block via --set-model (see nora agents). Mix freely — one Flow can use provider models and managed models in the same run.

​
Rate limits and quotas

Provider rate limits apply. If your Agents get 429s under load:
  • Bump your provider tier (OpenAI’s Tier 4, Anthropic’s Enterprise, etc.).
  • Add a fallback model per Agent (see nora agents update --set-model) — the Agent transparently switches on failure.
  • Route by intent — cheap intents to cheap models, hard intents to premium models.
Nora doesn’t add its own rate limits on top of the provider’s.

​
Cost dashboards

Provider costs show on the provider’s own dashboard (OpenAI, Anthropic, Google). Nora’s billing surface covers Nora’s platform costs; see Billing overview. Because keys are per-workspace, per-workspace cost attribution on the provider dashboard is straightforward.

​
Security

  • Keys are stored encrypted with a per-workspace key.
  • Rotation is instant.
  • Deletion is permanent (the key isn’t retained).
  • Every add / rotate / delete is audited.