← Beacon · embedded support
Each portal below is a real, separate tenant — its own project, signing secret and widget configuration in the database. Everything you can see between them — the shape and size of the launcher, how dense the panel is, what it is set in, how wide it opens, how its header and tabs read, which tabs exist at all, the tenant's own CSS, the pages it appears on — is server-side configuration an Admin changes without anyone touching the portal's code.
Open two of them and compare the panels. Northwind's is narrow, compact and monospaced because a trading desk reads numbers all day; Tailspin's is wide, round and roomy because it is selling holidays. Same widget, same source file, no fork — and none of it is colour.
Loading…
The six host pages are six different products — a trading blotter, a consignment note, a cloud console, a shop, a trip itinerary — and nothing on any of them changes when the widget mounts. It lives in a shadow root: their CSS cannot reach in and its styles cannot leak out, which is only demonstrated by pages that look nothing like each other or like us. And on Globex, try the checkout link — the widget is not hidden there, it is absent. That is a display rule, not a permission: the API answers on that page exactly as it does on any other.
Acme and Tailspin answer questions before they become tickets. Open either and try the Ask tab. Then ask something the help articles do not cover — where Tailspin flies, say, or Acme's refund policy. It declines and offers you a person rather than inventing an answer, and Talk to a person is on every screen of the conversation, including that one. That is the rule the whole feature is built around: the way out never depends on the assistant agreeing to it.
Not a language model. This showcase runs Beacon's built-in stand-in provider, which picks the sentences from the tenant's own published articles that best match the question, and cites the article it took them from. It is deterministic and offline, so a public demo needs no API key and costs nothing to leave running.
Everything around it is the real thing — retrieval, citation checking, the refusal when nothing matches, the turn limit, the escalation, the transcript that lands on the ticket. Only the sentence-writing is stood in for. In a real deployment that is Gemini, OpenAI or Anthropic, and Beacon refuses to let a tenant switch the assistant on until one of them is configured — a stand-in answering a real customer would be a confident answer that came from no model.