Support that answers from your actual docs — and knows when to get a human.
This page is the website of Copperline Coffee Co., a fictional six-person coffee roaster drowning in “where’s my order / how do I descale” emails. The chat bubble in the corner is BrewBot— a RAG agent trained on Copperline’s 15 help docs. Open it. Try to break it.
↘ It’s the copper button, bottom-right. No signup, no instructions needed.
<script src="https://brewbot.demo.intakecrew.com/widget.js" data-site="copperline" async></script>That’s it — one line, any platform, ~3 KB, no dependencies. This very page loads the widget the same way a client’s site would.
Try to break it
The interesting part of a support bot isn’t what it answers — it’s what it refuses to invent.
- →“Do you cater weddings?” — nowhere in the docs; watch it refuse and offer a human instead of guessing.
- →“Where is order CL-10482?” — live order lookup (canned demo data).
- →“Can I put my grinder burrs in the dishwasher?” — strong opinions, cited.
- →“What's your Bitcoin refund policy?” — invented-policy bait. It won't bite.
- →Ask the same thing 11 times — meet the friendly rate limit.
What’s under the hood
The honest pitch: Copperline is a demo scenario — a fictional business with hand-written docs. Everything else is real: the retrieval pipeline, the streaming answers, the citations, the refusal behavior, the handoff capture, the one-line install. This exact build is what IntakeCrew installs for real stores, trained on their real content, in days. Curious what it says about yourdocs? That’s the conversation we’d love to have.
Peek behind the curtain: conversations and handoffs land in /admin (basic auth — default dev credentials admin / copperline).