A comedy roast eval: 10 LLMs each tell one joke. You rate them 1–10. Highest score wins.

Models are called via OpenRouter; each gets a distinct ElevenLabs voice for delivery. Falls back to canned jokes and the browser's Web Speech API when keys aren't configured.

Rendering mermaid diagram...

File	Purpose
main.tsx	Hono server, routes, OpenRouter + ElevenLabs proxies
shared/models.ts	MODELS lineup, PROMPT, FAKE_JOKES, cleanJoke — shared between server and (via /api/lineup) client
frontend/index.html	Stage scaffold
frontend/script.ts	Client logic — fetches lineup, prepares show, plays audio, scores
frontend/style.css	Theme

Var	Effect when missing
OPENROUTER_API_KEY	Falls back to canned jokes from FAKE_JOKES
ELEVENLABS_API_KEY (or XI_API_KEY)	Falls back to browser TTS via SpeechSynthesisUtterance

• Test mode (default): canned jokes, no API calls to OpenRouter.
• Live mode: toggle the "live jokes" checkbox in the UI, or append ?live=1. Hits OpenRouter for real model jokes.

Method	Path	Returns
GET	/	The stage UI
GET	/frontend/*	Static assets (CSS, transpiled TS)
GET	/source	Redirect to this val's source on Val Town
GET	/api/lineup	{ prompt, models }
POST	/api/joke	{ joke, mode, model }
POST	/api/audio	audio/mpeg stream (or JSON error with browserVoiceFallback)

Click "view source" in the bottom-right of the live page, or fork from Val Town directly.