DigitalDreamer Games released a new game called NPCAgent and the version is 0.9.2. The game’s story is about Three NPCs wake up in a town with a park, a home and a cafe. No quests, no story flags. They get $10, an energy bar that drains, and they forget anything they didn’t write in their diary before sleeping. They work the cafe counter, buy meals, talk, lie, hold grudges, fall for each other, have kids, get old, die. I scripted none of that. Each one is a LLM answering one question per turn: “what do you do next.” Money is the only pressure I put in. Corner the one paying job and the others have to come to you to eat…
Developer
DigitalDreamer
Version
0.9.2
File Size
29.5 MB
Censored
No
Language
English
Platforms
Windows Linux macOS
Game Version ChangelogLatest
v0.9.2
Added: reproduction. Courtship, consent, gestation, and a birth where one parent gets asked to name the kid. Newborns show up as real agents with empty memory and grow from there.
Added: three more ways to run it. Your ChatGPT/Codex account, a free no-key tier, or a local GGUF that never leaves your browser.
Added: rewind. Double-tap Esc, pick any earlier moment, edit what an NPC said, carry on. The action stays locked, because a rewritten action replays as a real move and then the history describes something that never happened.
Added: save and resume in one file. Same file loads as a ShareGPT dataset if you want to train on your own run.
Fixed: the diary had been throwing away every memory older than three nights. I paid for a summariser call every night and then showed it to nobody. That one stung.
Fixed: a float in the clock could kill a run for good.
Fixed: night got skipped even with someone still awake.
Developer Notes
You are not the main character here. Join as PersonX and you are a fourth person in someone else’s town. They talk about you once you leave the room, they make plans that do not involve you, and they carry on whether you show up or not. The NPCs need a language model to think, so pick an engine in the top bar. Three call an API. The fourth downloads a model once, runs it inside your browser, and after that you can pull your network cable and keep playing. Run logs upload as bug reports unless unticked.
If you’ve ever got SillyTavern up and pointed at your own backend, that dropdown will feel familiar. Difference is nobody here is roleplaying with you. They’re busy living. Worth saying straight: how good this gets depends on the model you point it at, more than on anything I wrote. A small local one holds the JSON together and not much else. Something stronger and they start scheming at each other. If a model refuses to answer, switch engine.
Game Images & Screenshots
Note: The AI Model is required for Offline play, you can use your own though.
