[00] Legal
Hosted AI credits — data handling
What a hosted build sends, what is kept afterwards — your content is not, and the output only for a short redelivery window — and which subprocessors touch it on the way.
Last changed .
Why this page exists
The AI Content Builder can arrange your content two ways: with your own API key, or with credits hosted by us. The bring-your-own-key path sends your content to a provider you chose and have your own agreement with. The hosted path sends it to a provider we chose, through a service we run.
The second one needs explaining before anybody uses it, not after. A publishing plugin that quietly posts page content to an unnamed third party is the thing we would complain about in somebody else’s product.
How a hosted build runs
A build is not one long request. The plugin starts a job, our service checks your licence and credit balance and answers straight away with a job id, and the plugin then asks for the result every few seconds until the build has finished. The work itself runs in the same server process that accepted the request, after it has answered.
What is sent
When you build with hosted credits, the plugin sends our service:
- your licence key, in a request header, and the activated site URL and the site’s instance id, to authenticate the request and check the licence has credits;
- the content you pasted into the Build modal;
- the instructions the builder always sends, and the shape of your Flexible Content field — its layouts and sub-fields, as a JSON Schema — so the answer fits it;
- an output limit, and nothing you did not give it.
The service forwards the instructions, content and schema to [AI provider] (blank, to be supplied by the business), receives the arranged rows, and holds them for the plugin to collect. It does not add your email address, your account id, your other sites, your WordPress user list or anything else about you to what the provider receives.
What is retained
Nothing you send is kept beyond the request.
- The content you sent is not stored.
- The instructions and the field schema are not stored.
- None of it is written to a log, an error report, a trace or an analytics event.
It exists in the memory of the one process running your build, and is gone when that process ends. If the process is killed before it finishes, the build cannot be resumed — there is no copy to resume it from — so it is failed, your credits are returned, and the plugin asks you to try again.
The output, briefly
The arranged rows the provider sends back are the one piece of content that is written down, because the plugin has to come and collect them. They are stored once, in a table no account can read — only the service itself — and can be fetched again for a short redelivery window (ten minutes) in case a response is lost on its way to your site, so that a dropped connection does not cost you a second build.
After the window they are deleted: by the next request that asks for them, or by a scheduled sweep that runs every five minutes, whichever comes first. There is no transcript, no history feature, and no way for us to show you what you built last week. If a history feature is ever wanted, it is a new decision with a new version of this page.
What is kept
The metering has to be durable, so this part is:
| Kept | Why |
|---|---|
| Which licence and site ran a build, and when | Metering credits, per-licence rate limits, your dashboard |
| The provider, the model, and the tokens it reported | Pricing the build |
| How many credits were reserved and charged | The credit ledger your dashboard shows |
| Whether it succeeded, and an error code if not | Telling you why something failed, and finding our bugs |
None of those rows contains your content. An error is recorded as a code such as provider_timeout, never as the provider’s own message, because a provider’s message can repeat the text it was sent.
Credits are reserved before the provider is called, at the most the build could cost, and settled against what it actually used afterwards. A build that cannot be paid for is refused before anything is sent to a provider, and a failed build is not charged.
Subprocessors on this path
| Subprocessor | Role | Sees your content |
|---|---|---|
| [AI provider] (blank, to be supplied by the business) | Arranges the content | Yes, while the build runs |
| Vercel | Runs the service; request metadata and transport | In transit, and in the process running the build |
| Supabase | Holds the credit ledger, licence records and outputs | The arranged output, during the redelivery window |
The service can run on OpenRouter, OpenAI or Anthropic, and which one is a server setting rather than something in the plugin. The provider is the row that matters, and it stays blank until the production one is chosen, because the choice carries two commitments we are not willing to make vaguely — that the provider does not train on data submitted through its API, and that it retains submissions for no longer than the abuse-monitoring window its own terms describe. Those commitments go in this table, by name, with a link to the provider’s own terms, before a hosted credit can be spent. Where the provider offers it, the request asks it not to store the response: with OpenAI, every request is sent with storage turned off.
Account data rests in [data region] (blank, to be supplied by the business). A build is processed wherever the chosen provider processes it, which may be another region; that is a transfer, and it is named here once the provider is.
Bring your own key
Supplying your own provider API key is free, stays free, and keeps your content out of our infrastructure entirely: the plugin calls the provider directly and our service is not involved. There are no credits, no metering and no record here of what you built.
It is in the product for two reasons. It is the right answer for anyone with a procurement process or a data-processing agreement of their own, and it keeps us honest about what the hosted path is actually charging for — convenience and a bill we absorb, not access to a model.
What you can ask for
Because nothing you send is retained, and the output is gone within minutes, there is no copy of either to send you, correct or delete. What we hold and can act on is the metering above and your account record; the privacy policy covers both, and requests go to [address for privacy requests] (blank, to be supplied by the business).
If a provider holds a submission under its own abuse-monitoring retention, that is the provider’s record rather than ours, and this page will link to their process for it once the provider is named.
Before a hosted credit can be spent
- The production provider is chosen, named here, and its no-training and retention terms are linked.
- The data region and any transfer out of it are stated above.