Trip builder
Chat with an assistant about a trip, get a personal itinerary app for it.
Two agents:
- Chat interviews, thinks, and writes a brief. No web search, so it stays quick.
- Builder never talks to anyone. It takes one brief, researches the
destination properly, and produces the itinerary through custom tools.
The renderer in renderer/render.js is shared with
tools/itinerary-generator in the vault repo, and turns one itinerary JSON
into the same app shape as the Phu Quoc trip it was extracted from.
No storage anywhere
The itinerary is never written to a database. It is replayed from the builder
session's own event log, and the builder session id is recovered from the chat
session's log. The event log is the database.
People are separated by an anonymous session id in their own localStorage.
No accounts, no links.
Why polling
A build takes minutes, longer than a serverless function lives. /api/send
returns immediately and the client polls /api/state; each poll resolves
whatever tool calls are pending right now and returns. The chat keeps talking
while the builder is still researching.
Setup
Agents are persisted objects, created once:
node --env-file=.env setup/create-agent.js # first time only
node --env-file=.env setup/update-agent.js # to change a prompt
Ids live in lib/config.js. ANTHROPIC_API_KEY is read from the environment
if set, with a fallback in lib/config.js.
The agent's prompt and tool schemas live on Anthropic's servers, not in the
deploy. Editing lib/prompt.js, lib/editTools.js or any other tool schema
changes nothing anybody can see until update-agent.js runs. No build step
does it and a Vercel deploy will not. This has twice produced a change that
looks shipped, reads shipped in the diff, and is inert — so
GET /api/health?sources=1 reports chatAgent, which is either ok (vN) or
tells you exactly what has drifted and to run the script.
Tests
node --env-file=.env setup/test-builder.js # builder alone
node setup/test-ui.js # full flow through the real UI
