One buddy for every game. With your permission it sees this tab only, reacts to score changes, and talks back - same observe→reason→act engine as VibeCodeWorker. Turns meter Vibe Coins at true cost (25% cut included), live below + on /my/usage/.
Loading the game picker…
Talks while you play in any of 9 OpenAI voices (default Nova). Reads the screen, reacts to score changes, and runs on the VibeCodeWorker observe→reason→act loop. OpenAI + database turns meter Vibe Coins at true cost with the 25% cut included; every turn shows its Coins + CentiCentCoins below; local fallback answers are free.
👁️ Let the buddy see your screen (off unless you approve; the browser picker enforces your choice)
Tab only (safest): shares just this 4weird tab; other tabs, windows, and apps stay private. Full screen: you pick what to share in the browser picker; choose the game window to keep the rest private. Only one small snapshot per message is sent (never video, never stored).
🎙️ Talk to Buddy (optional mic; mute anytime, Buddy never records you)
Turn the mic on to interrupt Buddy mid-sentence; it stops, remembers what both of you said, and resumes from your cut-in. Only transcripts leave the device (never audio). Smart detection sends your sentence when you pause; turn it off to send manually.
📷 Let Buddy see you (optional camera; you approve every frame by sending)
Buddy reads your energy, posture, props, and backdrop to match your mood; kindly, never diagnosing, never identifying. While on, each message you send carries one small frame, itemized in that turn (~0.03 coins (3 centicentcoins) per frame). No video ever leaves the device, nothing is stored.
✨ Buddy avatar (optional 3D companion - 0.08 coins (8 centicentcoins)/min, free when hidden)
A cute three.js companion whose mouth follows Buddy's voice waveform and words, eyes blink, track your cursor, and widen with excitement. Presence meters by the minute while visible (this session's avatar: 0 centicentcoins).
Start a Buddy session to meet the avatar.
🎩 Wardrobe (one shop everywhere; every look costs 10 coins, looks-only, never pay-to-win)
Buying shows the exact price first and charges only what you confirm (singleplayer boosts live behind guarded dev charges, capped at 10,000 coins each and 1,000/day per game; multiplayer boosts are banned).
hat
glasses
outfit
accessory
effect
⚡ Buddy extras (streaming answers are automatic; memory + auto-reacts are opt-in)
Buddy is off. Start a session to talk while you play.
Avatar, TTS voice, and personality live in the session widget above (no page scroll needed) — pick a voice below to preview free, then set it in the widget when you start. Component-owned meters (waveform, token burn) render in-widget; queued as handoff, not touched here.
Preview any voice free from the widget - no session, no coins.
Models: tts-1 (fast, $15/1M chars) + tts-1-hd (rich, $30/1M chars). Speeds 0.5x-2.0x. Chat runs on BUDDY_MODEL (default gpt-4o-mini). Every turn is metered at true upstream cost plus database writes, converted to gross coins with the 25% cut included and shown per turn in Coins + CentiCentCoins (100 centicentcoins = 1 coin).
| What | Unit | Gross | Notes |
|---|---|---|---|
| Chat turn | tokens + snapshot | ~0.01+ coins | Prompt + reply + optional 1k-token image + DB leg |
| Voice reply | chars | ~0.01+ coins | tts-1 vs tts-1-hd rate + DB leg; browser speech is free |
| 3D avatar | minute | 8 centicentcoins | Only while visible; hidden = free |
| Camera frame | frame | 3 centicentcoins | One frame per message you send, never silent |
| Local fallback | reply | free | When AI is unavailable; browser speech included |
Wallet running low? Turns fail fast with a 402 before any AI spend. Check /my/usage/ →
The buddy works from game text + score events without any permission. If you want it to actually see the game, start a session and press Share this tab only - the browser picker is constrained to this 4weird tab, so other tabs, windows, and apps stay private. At most one small JPEG snapshot per message is sent to the model (never video, never stored) - unchanged screens skip the re-upload so the turn answers faster, and when screen + camera ride one turn the model sees both; stop sharing anytime from the widget or the browser bar. Snapshot turns carry a small image-token charge, itemized in that turn's cost line.