Your key, your bill
AnswerRails does not sell you AI. You bring an account with an AI provider, AnswerRails connects your site to it, and the provider bills you directly for what your visitors use. There is no platform in between.
How bring-your-own-key works
You open an account with one of the supported providers, create an API key, and paste it into AnswerRails’ settings. From then on, every answer your assistant gives is a call from your WordPress server to that provider, authenticated with your key, and billed to your account at the provider’s own rates. AnswerRails adds nothing to the per-message cost and never sees the invoice.
Three providers are supported on every plan, free and Pro alike:
| Provider | Notes |
|---|---|
| Mistral | The default for new installs. An API key is free to create at console.mistral.ai with no card, and Mistral offers a zero-cost tier — which is why a new install needs no OpenAI account. The setup walkthrough shows you how to turn off Mistral’s free-tier training before your first answer. Some Mistral models are labelled in the plugin as needing Mistral’s pay-as-you-go tier; that is a statement about Mistral’s tiers, not ours. |
| OpenAI | Chat via the Responses API with automatic failover to Chat Completions. An OpenAI key also unlocks the optional vector store for knowledge search (below). |
| Anthropic | Claude models. Moderation on this provider runs as a classifier prompt on the Claude model you configured, so it is a small call on your own bill. |
Provider and model are your choice on every plan; nothing is locked to a tier. Each provider’s catalogue is curated, and there is an escape hatch for a model id that isn’t listed yet — with the honesty that the plugin cannot vouch for a model it doesn’t know. When a provider announces a model’s retirement, the admin shows you an advisory with a suggested replacement: a snoozable warning up to 60 days out, a non-dismissible one inside 14.
Who is billed for what
- Answers. Each visitor question is one chat request to your provider, containing the question, recent conversation context, the relevant excerpts from your knowledge base and the list of your site’s page titles. The provider bills you per token, at its published rates.
- Moderation. Every visitor message is checked by the same provider before it is answered. OpenAI’s moderation endpoint is free; Mistral’s is priced per token at a fraction of a cent per message; on Anthropic it is a short classifier call on your chosen model. One vendor relationship per site: AnswerRails never makes a cross-provider call and never needs a second key.
- Knowledge search. By default your knowledge base is searched locally, in your own database, at no cost. If you have an OpenAI key and choose to create a vector store, your knowledge items are also uploaded to a vector store on your OpenAI account, and OpenAI’s storage and search charges apply there.
- The plugin. The free plugin costs nothing. Pro is a flat-priced licence, unrelated to how many questions your visitors ask. Pricing has the figures.
What AnswerRails itself never does
No AnswerRails or Celtic3D server is in the path of a conversation. Your visitor’s message goes from their browser to your WordPress site, and from your site to your provider. Nothing is sent to any AI provider until you have entered a key, and the only provider contacted is the one you configured. Licence checks for Pro read a locally cached state, so there is no network call on the chat path and a licensing outage never interrupts a paying site mid-conversation.
We chose this deliberately over proxying your traffic through a service of ours. A proxy would have let us sell you AI by the message; it would also have put us in the middle of your visitors’ conversations, which is the opposite of the product we wanted to build.
Where your key lives
Your key is stored in your WordPress database, as a plugin option, and is sent only to the provider it belongs to. It is never written into the page your visitors load; nothing secret reaches the browser. On uninstall, keys are always deleted, whether or not you chose to keep the rest of the plugin’s data.
Testing the connection
The Overview screen has a one-click Test APIs probe. It makes a real call to the chat endpoint, a real search against your knowledge base (vector or local, whichever is in use), and a real moderation check on a fixed benign string, and records the results as health evidence. A test call is billed like any other call. The same screen shows rolling success rates and latency per service, the last error seen, and a plain-language reading of what it means and what to do.
