Skip to main content
Every company that adopts Vendo has to demo it: to its own executives, to a prospect, to a new user on day one. A live agent is the wrong thing to put in front of an audience — it is slow, it is different every time, and the one run that matters is the run where it improvises. Tour mode is the fix. You record the handful of answers the demo depends on, and Vendo replays them at a live turn’s cadence. Everything else — every improvised question, every follow-up about what is on screen — reaches the real agent, untouched.
Ask that question and the answer streams back with no model call. Ask anything else and nothing changed.

​
Matching is strict, and once

Two rules keep a tour from swallowing the demo it is supposed to carry. An entry fires only on a close variant of its own frozen prompt. Matching is a normalized similarity score, not keyword presence — so a typo, a dropped em dash, or a small word swap still lands the entry, while a different ask about the same subject does not:
That last pair is the whole reason the rule exists. Keyword matching cannot tell “ask for this” from “change the thing you just made”, because both sentences are about rent being late — so it replays the recording on top of the app the audience just watched arrive. An entry fires at most once per thread. After it plays, the same sentence falls through to the live agent. This is what makes a follow-up an edit of what is on screen rather than a second copy of it. The state is derived from the thread’s own transcript, so it survives the live turns in between.

​
What an entry can replay

respond takes prose, a recorded app, or a sequence of both.
prompt takes an array when the same answer has more than one phrasing — the line you say out loud and the suggestion chip beside it are the usual two. An app part is a real app, not painted pixels: the document is imported as an owned copy for the signed-in user, so it opens, pins, survives a reload, and — the point — can be edited by the next turn, which is the live agent’s. buildMs is how long the build takes (default 8 seconds). A real generation of a dashboard-sized app takes 30–60 seconds; raise buildMs toward that when the audience is meant to believe the app was generated in front of them.

​
Recording an app

Build the app once through the real agent, then export it:
Its app.json is the document — drop it in your repo and import it:
Any id in the document is ignored: each replay mints a fresh one, exactly as a real generation would.

​
Determinism

A rehearsed demo has to run the same way twice, so every pause and every chunk boundary is drawn from a stream seeded by the entry’s own prompt — uneven like a live provider, and the same unevenness on every run. Nothing in a tour calls Math.random. Pacing is measured against real turns rather than invented: a delta carries 20–55 characters and lands 76–530ms after the last, with the odd burst and the odd stall. Marks are absolute offsets from the moment the turn opened, so a slow store spends its latency inside the budget instead of on top of it — the same tour reads the same way on a laptop and on a hosted deployment.

​
What tour mode is not

  • Not a key-gated feature. Tours are plain config with no Cloud dependency: identical behavior with and without VENDO_API_KEY.
  • Not a mock. A tour turn streams the same wire parts a live turn streams, persists into the same thread, and leaves a real app behind.
  • Not a fallback. If a replay cannot run, the turn fails loudly and names the entry in the server log. A tour that silently degrades is one you find out about on stage.