Docs / Set up your org

Configuring the AI provider key

Every analysis and canvas REEZN generates runs on an AI provider you configure once, at the org level, under Org settings → AI. There is no per-project or per-member key: one provider, one key, and one set of model choices power generation for the whole organization.

You set this up either in the onboarding wizard when you first create the org, or any time afterward from Org settings → AI. Without a working provider and key in place, there is nothing to generate from, so this is one of the few things that has to be done before features can move past draft.

The three steps

Choose a provider. Pick the one whose key you hold. REEZN supports several directly, or you can route everything through the Vercel AI Gateway with a single key.
Enter the API key for that provider. The field is masked as you type, and once you save it is never shown back to you; it becomes write-only.
Pick a model for each task. Three stages each take their own model: analysis generation, canvas generation, and section regeneration. Every provider offers a few options tuned for quality, balance, or speed and cost.

Supported providers

You choose one provider for the whole org. Seven are supported directly, plus the Vercel AI Gateway as a single front door to all of them:

Vercel AI Gateway: A single key that routes to any underlying provider, handy when you’d rather manage one credential and switch models without juggling separate keys.
Anthropic: The Claude family, spanning a highest-quality, a balanced, and a fast-and-cheap tier. Claude is the recommended default.
OpenAI: The GPT and reasoning-model families, across the same quality, balance, and cost tiers.
Google: The Gemini family, from its highest-quality models down to the fast, low-cost Flash tier.
Mistral AI: The Mistral family, from its large flagship models to small, fast ones.
Groq: Fast inference of open models, for teams optimizing for speed and cost.
xAI: The Grok family, across quality, balance, and cost tiers.
DeepSeek: DeepSeek’s chat and reasoning models, the latter for reasoning-heavy work.

Each provider’s current model lineup is shown in the dropdown when you select it, labelled by what it is best at, so you always pick from what that provider offers today rather than a list that ages in the docs.

Switching providers reselects sensible default models for that provider, so if you change from, say, Anthropic to OpenAI, the three task models reset to OpenAI’s defaults rather than carrying over model IDs that provider doesn’t recognize. Review the model choices after any provider change.

Entering the key: write-only after save

The API key field is masked while you type, and the provider sets the placeholder so you can see the shape of key it expects (for example sk-ant-… for Anthropic, sk-… for OpenAI). When you save, the key is stored on the org and never displayed again; the field comes back empty or masked, not pre-filled with the secret.

To rotate or replace a key, just type the new one and save again; you don’t need (and can’t) read the old one back first. This write-only handling means a shared screen or a screenshot of the settings page never leaks the credential.

Whenever you save, REEZN checks the key against the chosen provider before writing it; an invalid or expired key is rejected with an inline error and the save is blocked, so a typo doesn’t silently break generation later.

Model per task

Model selection isn’t a single choice. REEZN lets you set a model independently for each of three stages, all drawn from your chosen provider’s lineup:

Analysis generation: The model that drafts the analysis document, the artifact that frames everything downstream. Worth spending on quality here.
Canvas generation: The model that drafts each service’s REASONS canvas from the approved analysis. A balanced model usually fits.
Section regeneration: The model reserved for regenerating an individual section. Defaults to the provider’s fastest, cheapest option, since such passes are quick and targeted.

The three tasks have different demands, so setting them independently lets you spend where it matters: a top-tier model for the analysis that frames the work, something balanced for the canvases, and a fast, cheap model for short, targeted regeneration passes.

Why this exists

AI config lives at the org level so every plan is generated with the same provider, key, and model choices, a consistency you’d lose if each member brought their own credentials. Keys are write-only after save so a leaked screen never exposes a secret. And splitting the model choice across the three tasks means you’re not forced to pay for a flagship model on a quick regeneration, or to skimp on the analysis that everything else is built from.

AI configuration is admin-only. The whole AI tab, and every other org setting, is only reachable by an org admin; non-admins see a note asking them to contact one. This is enforced server-side, not just hidden in the interface.