MarketFish is a multi-agent market simulation engine: instead of asking one LLM "will this product succeed?", it builds a digital market with 128+ AI consumers - each with its own identity, budget, emotions, and biases - and lets them shop across 30 rounds. Their purchase decisions, churn patterns, and social influence reveal what real users would do, built on six academic papers (Generative Agents, OASIS, TwinMarket, Agent Bazaar, EconSimulacra, SMIF) and 11 LLM providers.

Use it in one of three modes: Explore (feed it seed data and let AI discover and rank product directions), Validate (test your own product idea for a survival score, buyer profiles, and optimal price), or Hybrid (pit your idea against AI-generated competitors in the same sandbox). Compared to MiroFish, the most well-known general-purpose social-simulation engine, MarketFish is scoped specifically to product-market prediction, runs as a single Streamlit app with zero external memory dependencies (versus Flask+Node.js+Docker and an external Zep Cloud memory service), and supports 11 LLM providers across China and global markets rather than OpenAI-compatible only.

Input: seed market data (static JSON) or your own product idea and pricing. Output: a five-stage pipeline result - market ontology, a knowledge graph of entities and pain points, 128 heterogeneous AI consumer agents, a 30-round simulation, and an evidence report of who bought, why, and what killed competitors.

git clone https://github.com/Key-wxh/market-fish.git
cd market-fish
cp .env.example .env
pip install -r requirements.txt
streamlit run streamlit_app.py

The CLI runs the same pipeline headlessly: python run.py --mode validate --name "My App" --pricing "$10", with --reuse-agents to reuse a prior agent population and save cost.

Six V6 modules implement the underlying papers: Memory (Generative Agents, UIST 2023) lets agents remember purchases and reflections; a Time Engine and RecSys (both from OASIS, 2025) add realistic 24-hour activation and personalized recommendations; BDI v2 (TwinMarket, NeurIPS 2025) runs a six-step cognitive loop with behavioral biases; Stress (EconSimulacra, 2026) adjusts willingness to pay under financial or social pressure; and Grounding (SMIF, ETASR 2026) applies RAG plus rule constraints for realistic decisions. Supports 11 LLM providers - DeepSeek, Qwen, Doubao, Zhipu, Baidu, and Hunyuan in China, plus OpenAI, Anthropic, Google, Mistral, and Meta globally - with just one API key required to run.

Founders and product teams who want simulated market evidence - buyer profiles, pricing, and competitive survival odds - before launching, instead of guessing or asking a single LLM for a verdict. Open source under the MIT license.