Send Hold

Send Hold is an agentic send-gate tool for live-night folios. Night-of talent ops (and the agents that assist them) draft a plan: Support, Dinner, Drink around a headliner. A generic agent would send that draft. Send Hold does not — it calls Qloo, checks affiliation, and only allows send when every line is in the headliner's affiliated set.

Who it is for: the person running the night — talent ops, production, or venue ops — who cannot afford to fire the wrong dinner or drink next to the headliner.

Why Qloo is essential: Without the taste graph, nothing can know whether a dinner or drink belongs with a given headliner. Hold / clear / silent comes from Qloo /search and /v2/insights affiliation checks — not from an LLM guessing taste.

Generic vs Qloo (see the two panels on the demo):

• Left: Generic agent would send — the draft as written.
• Right: After the hold — Qloo result; out-of-set lines are held with an in-set alternative when available.

Agent tool (official hackathon shape: agentic tool):

• GET https://send-hold.vercel.app/api/agent — tool descriptors and curl examples
• POST /api/agent/evaluate — { "planId": "north-wharf" } or custom headliner + three lines
• POST /api/agent/send — re-evaluates; refuses held/silent; logs when cleared

Pipeline (deterministic, server-side):

1. Resolve headliner and each line via Qloo search (artist / place / brand).
2. Check each line against the headliner's affiliated set via insights.
3. Gate send: held or silent → block; cleared → server re-evaluates before any outbound action.
4. When held, suggest an in-set alternative from affiliated insights.

The hosted demo at https://send-hold.vercel.app/ walks four scripted folios (held, held, silent negative control, cleared) in about 60 seconds. Expand Agent tool on the sheet for live JSON + curl. Signed-in product at /app (Convex Auth, custom folios, history) uses the same evaluation core.

Stack: Next.js, Qloo API (hackathon.api.qloo.com), Vercel. Product path: Convex, optional Resend email.

Honest scope: This is not an autonomous LLM chatbot. Qloo is the permission check; agents call evaluate/send as tools.

What to try (60 seconds)

1. North Wharf — generic panel shows Joe Pizza; Qloo held, Cannot send
2. Canal Room — drink held (brand out of set), Cannot send
3. Late Set — no answer (negative control headliner)
4. Open Door — cleared → Send plan → Send log entry
5. Open Agent tool — copy curl for POST /api/agent/evaluate on the current folio

Banner should read Live Qloo evaluation when the host has QLOO_API_KEY set.

Built With

• convex
• nextjs
• qloo
• typescript
• vercel