Generation

Litestore wires three AI providers through one shared integration layer (the AI SDK): Google Gemini, Anthropic Claude and OpenAI.

You pick one, supply your own key, and a small set of admin features start working:

• Meta title and description
• Tag suggestions
• Category suggestions
• Attribute discovery
• Email and social drafts
• Product import from a URL
• A task assistant

Every one of them prefills something for you to review. None of them writes to the storefront on its own.

That is the governing rule of this page, and the rest of it is mostly an explanation of how the rule is enforced rather than merely intended.

Choosing a provider

One provider is active at a time. Each has a default model and its own key:

Google (default)   gemini-2.0-flash           GOOGLE_GENERATIVE_AI_API_KEY
Anthropic          claude-sonnet-4-20250514   ANTHROPIC_API_KEY
OpenAI             gpt-4o                     OPENAI_API_KEY

You choose with AI_PROVIDER and supply the matching key.

AI_MODEL overrides the model name if you want a different one from the same provider.

Google is the default, so a store that sets only GOOGLE_GENERATIVE_AI_API_KEY needs no further configuration.

Failover

If your provider is rate-limited or returns a server error, the request retries on the next provider in the chain:

Google → Anthropic → OpenAI → Google

It only moves to a provider whose key is also set. Failover is not a way to get a second provider for free.

There is one deliberate exception. A rejected or revoked key never falls back. A broken credential surfaces as a failure instead of being masked by a silent hop to another provider — otherwise you would discover months later that a key you thought was in use had been dead the whole time, and that your bill was somewhere else.

Set AI_FALLBACK_ENABLED to false to switch the chain off entirely.

Where the buttons are

Four places in the admin carry a generate control:

• SEO meta title and description, on the product, collection, category and page forms
• Category suggestions, on the product's category card
• Variant attribute discovery, in the specifications editor
• Email subject and body, in the template and compose forms

Endpoints without a button

Two capabilities ship as working endpoints that no admin control calls:

• Tag suggestion
• Social post drafting

Settings → AI lists tag generation as used by the external API. You reach both through the API, not through the admin.

This is worth stating plainly because "it exists" and "there is a button for it" are different claims, and the difference is exactly what people get wrong when they plan work around a feature list.

Product import from a URL

With a provider configured, you can paste a product URL and have its content extracted into a draft.

Without one, that import option is disabled in the dialog rather than failing when you use it.

Nothing publishes itself

The SEO generator writes its result into the form you are editing, and does nothing else.

You read it, change what is wrong, and save. There is no auto-publish path — not a setting for one, and not a hidden one.

The assistant proposes, you approve

The task assistant can read store data directly. It cannot make a change.

Any change-making tool returns a pending proposal plus a server-signed approval token. The change runs only after you click Approve.

you ask
    ↓
assistant reads store data directly
    ↓
change-making tool returns a proposal + signed token
    ↓
you click Approve
    ↓
the change runs

Money, refunds and stock go further: they are proposal-only by construction. An execute-mode tool on those permissions fails to build. This is not a runtime check that a future change could weaken by accident — code that would let the assistant move money directly does not compile.

Settings → AI records every proposal, and whether it was approved, dismissed or expired.

The autonomy ceiling

Settings → AI carries a store-wide autonomy ceiling. It defaults to read-only.

Alongside it are individual capabilities you can switch off.

A capability above the ceiling, or switched off, is never shown to the model at all. It is not offered and then refused — the model does not know it exists, so there is nothing to talk it into.

Input screening

Prompts are checked for injection patterns before they reach a provider.

A blocked message is refused rather than forwarded.

What ships, and what does not

Getting started

Add one provider key to your environment, and set AI_PROVIDER if it is not Google.

Open Settings → AI and use the test button to confirm the connection. The panel names the active provider, the model and the fallback.

Then decide what the assistant may do. Leave the ceiling at read-only, or raise it and switch off the individual capabilities you do not want — remembering that anything above the ceiling is never offered to the model in the first place.

Generate where the buttons appear — the SEO card, the category card, the specifications editor, the email composer — then review and save.

The useful question before turning any of this on is not "can the AI do this?" It is:

What would I have to check before I would be willing to publish this unread — and since I would always check it, is read-and-approve actually costing me anything?

Related