What it does MiroFish ingests “seed” documents—news, policy drafts, financial signals, even novels—and spins up a parallel digital society. Thousands of LLM-backed agents with distinct personas, memories, and social ties interact and evolve. You can watch, intervene, or just ask for a forecast. The output is a prediction report plus a living world you can still chat with.

The interesting bit The project explicitly bills itself as a rehearsal lab for decision-makers and a sandbox for creative what-ifs. That dual use—Wuhan University public-opinion crises on one hand, lost endings of Dream of the Red Chamber on the other—is unusual. The simulation engine itself is borrowed from the open-source OASIS project; MiroFish adds the prediction-reporting layer and a web UI on top.

Key highlights

• Workflow: GraphRAG knowledge graph → persona generation → dual-platform simulation → auto-generated report → post-simulation chat with any agent.
• Requires an OpenAI-compatible LLM API (docs suggest Alibaba Qwen-plus) and a Zep Cloud key for memory.
• Docker Compose or local Node/Python stack; frontend on :3000, backend on :5001.
• Live demo site and Bilibili videos of public-opinion and literary simulations are linked.
• Incubated by Shanda Group; built atop CAMEL-AI’s OASIS engine.

Caveats

• The README warns of “high consumption” and suggests starting with fewer than 40 simulation rounds—costs add up fast.
• Financial and political prediction demos are listed as “coming soon,” so those use cases remain unproven in the open repo.

Verdict Worth a look if you’re experimenting with multi-agent social simulation or need a structured way to stress-test narratives. Skip it if you want battle-tested forecasting; this is a research-grade sandbox, not a crystal ball.