Give every NPC a real mind, backed by your own lore.
GameRagKit is a retrieval-augmented persona engine for game NPCs: drop your world,
region, and faction lore into files, declare a persona in YAML, and get an NPC that
answers in character, remembers what it's told, and can share personality context
with hundreds of others through the same hierarchy, all through one small HTTP
API (/ask) any engine or client can call.
See it running in a real scene
Four playable Three.js first-person scenes, each backed by real GameRagKit NPCs,
not scripted dialogue. Walk up, press E, and every reply comes
from an actual /ask call.
Single NPC
One guard in a small courtyard, the minimal proof that a real NPC works inside a first-person game scene.
1 NPCCity
Five NPCs with distinct personas around a town square, proving concurrent NPCs don't cross-talk with each other.
5 NPCsMetropolis
Hundreds of clickable crowd figures, each sharing one of 14 archetype NPCs: the cluster pattern that keeps backend load bounded at any crowd size.
900 figures · 14 shared NPCsTavern Table (voice)
Talk out loud to three NPCs at once. A router picks who answers, they reply to each other by name, and each one has its own local voice.
3 NPCs · local STT + TTSHear what a GPU would add
The NPCs in these demos speak with Kokoro, a small open voice model that runs on the CPU server hosting this site. The same lines below, generated with two larger open models, carry real emotion: laughs, sighs, voices designed from a one-line character description. They need a GPU to keep up in a live conversation, and we can’t afford to host one for this demo. On a gaming PC with a modern GPU, GameRagKit can run them locally, and your NPCs would sound like this.
Oh, you should have seen his face! Bram swore up and down he’d paid his tab, so I just held up the ledger. Ha!
Busy? I’ve three swords to finish before sundown, and she wants to squabble over a few coppers. Unbelievable.
Hear ye, hear ye! A royal messenger rode through Riverside last night, and he left something behind at the tavern!
All three are open models you can ship (Apache-2.0 / MIT).
Chatterbox clones the Qwen-designed voices and adds [laugh] and
[sigh].
Help us put a GPU behind these NPCs
Hosting, LLM usage and every voice in these demos are paid out of pocket. Contributions go straight to running them, and to a GPU server so everyone visiting can hear the expressive voices above instead of the CPU ones.
RedotPay ID: 1236302071