What is this?
Wispverse (芥子) is an AI-native evolution simulation game: 15-20 anthropomorphized characters autonomously develop division of labor, society and culture in a garden, while you play the 21st character — a disguised infiltrator who secretly steers the civilization through rare decision moments until they eventually identify you.
Wispverse (芥子) is an AI-native evolution simulation game: 15-20 anthropomorphized characters autonomously develop division of labor, society and culture in a garden, while you play the 21st character — a disguised infiltrator who secretly steers the civilization through rare decision moments until they eventually identify you.Is this an actual playable game, or more of an AI agents demo?A playable web game: the garden is rendered on Canvas 2D by a React 18 + TypeScript + Vite frontend, backed by a Python FastAPI server where every character runs its own state machine driven by LLM decisions.
A playable web game: the garden is rendered on Canvas 2D by a React 18 + TypeScript + Vite frontend, backed by a Python FastAPI server where every character runs its own state machine driven by LLM decisions.How do I get it running?There is a zero-dependency demo mode (in-memory storage plus a rule-template LLM) where you just open the browser and enter the garden as one of 21 identities. Production mode needs a running local PostgreSQL (127.0.0.1:5432) and an LLM_API_KEY; the script then creates the wispverse database, injects the env vars and starts uvicorn.
There is a zero-dependency demo mode (in-memory storage plus a rule-template LLM) where you just open the browser and enter the garden as one of 21 identities. Production mode needs a running local PostgreSQL (127.0.0.1:5432) and an LLM_API_KEY; the script then creates the wispverse database, injects the env vars and starts uvicorn.Do all those NPC thoughts burn expensive LLM calls?No — daily NPC behavior runs on a rule engine with zero LLM calls. An LLM gateway (opencode + deepseek-v4-flash) handles decisions with cost-tiered routing and failure fallback to rule templates, and embeddings (bge-m3) run locally. Production mode does require an LLM API key and PostgreSQL with pgvector.
No — daily NPC behavior runs on a rule engine with zero LLM calls. An LLM gateway (opencode + deepseek-v4-flash) handles decisions with cost-tiered routing and failure fallback to rule templates, and embeddings (bge-m3) run locally. Production mode does require an LLM API key and PostgreSQL with pgvector.Can I pick and name my own character?Yes — creation is a single-page flow with 5 avatars (Leaf Shadow, Creek Spirit, Ink Heart, Forest Song, Cloud Wanderer) plus a surprise easter egg, and the 21st character enters the garden under your custom name, falling back to the avatar's default.
Yes — creation is a single-page flow with 5 avatars (Leaf Shadow, Creek Spirit, Ink Heart, Forest Song, Cloud Wanderer) plus a surprise easter egg, and the 21st character enters the garden under your custom name, falling back to the avatar's default.Is there anything for safety, like parental controls?Yes — a compliance middleware ships with a minors mode, AI-content labeling and session-length monitoring that nudges you in-world after 2+ hours. Input also passes a three-gate poisoning filter with audit rollback to block prompt-injection attempts.
Yes — a compliance middleware ships with a minors mode, AI-content labeling and session-length monitoring that nudges you in-world after 2+ hours. Input also passes a three-gate poisoning filter with audit rollback to block prompt-injection attempts.How well tested is all this agentic machinery?Heavily: 625 pytest cases across 44 modules cover agent state machines, memory, exposure propagation, era transitions and real PostgreSQL reads/writes, plus 11 Playwright E2E specs with 43 cases spanning onboarding, decision moments and the tool panel.
Heavily: 625 pytest cases across 44 modules cover agent state machines, memory, exposure propagation, era transitions and real PostgreSQL reads/writes, plus 11 Playwright E2E specs with 43 cases spanning onboarding, decision moments and the tool panel.What are similar projects?MemoryBear, Octop, oto-dock, agentscope, gemini-fullstack-langgraph-quickstart, openhikmah-web, printfilm, m_flow
MemoryBear, Octop, oto-dock, agentscope, gemini-fullstack-langgraph-quickstart, openhikmah-web, printfilm, m_flowWhich topics is it in?pgvector, llm-gateway, simulation, fastapi, game-development, agent-memory, memory, game, vite, rag, deepseek, postgresql
pgvector, llm-gateway, simulation, fastapi, game-development, agent-memory, memory, game, vite, rag, deepseek, postgresql