Helpie’s assistant runs on Claude, made by Anthropic — and it runs on your Anthropic account, not ours. You bring your own API key, and every AI call goes directly from your site to Anthropic. Helpie never proxies your customer conversations through a third party, and there’s no middleman marking up your usage.
This page walks through connecting a key, choosing a model, telling Helpie what your site does, and — importantly — what all of this costs and how to keep that under control. (New to Helpie entirely? Getting started covers the plugin as a whole; this page is just the AI half.)
—
Getting and Saving Your Key
- Create a key at console.anthropic.com (you’ll need an Anthropic account with billing set up)
- Go to Helpie > Settings > AI
- Paste the key into the API key field and click Save AI settings
- Click Test API key — this sends a tiny “ping” message to Anthropic and tells you plainly whether the key works
The key is stored encrypted in your WordPress database and is never echoed back in full — the field only ever shows the last four characters. If you need to replace it, paste a new one; if you need to remove it, tick Clear saved key (the AI stops working until a new key is saved).
If the test fails, the error message tells you which kind of failure it was — a rejected key, a rate limit, or your server being unable to reach api.anthropic.com — with the full response available to expand if you need to dig deeper.
—
Choosing a Model
The Model dropdown offers three Claude models: Haiku 4.5, Sonnet 4.6, and Opus 4.7. They trade off cost against reasoning ability — Haiku is the cheapest and fastest, Opus the most capable and priciest, with roughly a twenty-fold price difference between the two ends.
Haiku is the recommended choice for most chat traffic, and it’s the default. Customer support questions are mostly retrieval — “find the right page and summarise it” — which Haiku does well. Try Haiku first; move up only if you find its answers genuinely lacking.
—
Telling Helpie What Your Site Does
Two fields on the AI tab give the assistant context it can’t get from individual pages:
What does this site do? — a short description of your business, in your own words. Helpie uses this to judge whether a question is even in scope (someone asking a hoodie shop about the weather gets a polite redirect, not a hallucinated forecast). There’s an Auto-detect from indexed content button that reads a sample of your indexed pages and asks Claude to draft this for you — you review and edit before saving. The auto-detect uses one small AI call.
Always in context — content the assistant should always have on hand regardless of which page a question matches. Footers and navigation menus often don’t appear in indexed page bodies, so things like your full product list, “where to buy” link, or support contact can slip through the cracks. Paste them here and they’re always available.
—
How Grounding Works, in Plain Words
The assistant follows one hard rule: it only answers from your indexed site content, and it cites where each answer came from.
When a visitor asks something, Helpie searches your knowledge base, hands the best-matching excerpts to Claude, and instructs it to answer from those excerpts alone — general knowledge from the model’s training is explicitly off-limits, because Helpie has no way to know whether that knowledge matches your site. Every factual claim must point back to a real source, and those sources appear as links under the answer.
When your content doesn’t cover a question, the assistant doesn’t guess. It asks one clarifying question, and if that doesn’t get there, it offers to pass the question to your team as a ticket. An answer it can’t back up is an answer it won’t give — which means the chat can be wrong about far fewer things than a free-wheeling chatbot, at the cost of occasionally saying “let me hand you to a human.” We think that’s the right trade for a support tool.
Safety refusals work the same way. The assistant declines to give legal, medical, or financial advice — always, by design — and routes those conversations to a person instead. It can still point a visitor at your published content on those topics (“here’s our returns policy page”), it just won’t opine on how the law or medicine applies to their situation. That’s a routing feature: the questions that carry real consequences land with you, not with an AI.
You can layer your own personality and deflection rules on top — see The Assistant’s Persona.
—
What It Costs
Honesty first: every AI-generated chat answer uses your Anthropic key and costs a small amount of money. So does a clarifying question, the Auto-detect button above, and each Content Prospector refresh. Things that cost nothing: the opening greeting (it’s static text), browsing the widget, indexing your content, and route-rule matching.
Individual answers on Haiku cost fractions of a cent, so a typical small store’s monthly chat spend is modest — but “modest” should be your decision, not a surprise. Which is why spend caps exist:
- Go to Helpie > Settings > AI Spend
- Set daily, weekly, and monthly caps — even a tiny daily cap protects you completely while you’re testing
- Watch the progress bars: today, this week, this month, each against its cap
- Choose what happens when a cap is hit: the widget shows a ticket form, a direct-contact form, or hides until the cap resets
Costs shown in Helpie are estimated from token usage; your actual Anthropic invoice is the source of truth.
—
Trying It Before Anyone Sees It
The chat widget feature toggle ships off, and while it’s off the AI Tester (Helpie > Settings > AI Tester) gives you the full chat experience privately — same grounding, same persona, same everything, plus a debug readout of what each answer cost and which sources it retrieved. Calibrate there, then publish. The chat widget page covers going live.
—
Common Questions
Does Anthropic see my customers’ questions?
Questions are sent to Anthropic’s API to generate answers — that’s how the AI works — using your own account and their API terms. No one else sits in the middle. For what Helpie stores locally and for how long, see Privacy.
What happens if my key stops working or my Anthropic credit runs out?
The assistant stops being able to answer, and the widget falls back to offering the ticket form — visitors can still reach you. Fix the key or billing at console.anthropic.com and it picks straight back up.
Can I change models later?
Any time. It’s a dropdown — the change applies to the next message sent.