What it is
An open-source (AGPL-3.0) multi-agent swarm-simulation app that builds a parallel “digital world” from seed materials and outputs a prediction report.
Gabriel’s notes
MiroFish is an open-source “AI prediction engine” that takes real-world seed materials (e.g., news, policy drafts, financial signals, reports, or even fiction), builds a parallel simulation environment, and produces a prediction report plus an interactive simulated world. The repo positions this as a multi-agent / swarm-intelligence approach where large numbers of agents interact and “socially evolve.” Accuracy claims like “predict anything” are a project tagline, not something I can independently validate from the public materials. Unknown / not confirmed.
Quick take: This is a big, ambitious multi-agent simulation repo with a glossy “SimCity for forecasting” vibe. It might be legitimately useful—just assume the first thing it will predict is your token bill.
I saved this under AI because it’s basically a wrapper around “let a crowd of agents argue inside a constructed world” as a forecasting workflow—i.e., the most 2026 way possible to turn uncertainty into structured output.
Good fit if you want to:
- Turn messy seed material into a structured simulation + report (instead of a single LLM answer).
- Stress-test “what if” decisions (policy, comms, product moves, market narratives) via scenario injection.
- Prototype multi-agent social simulation UX (front-end + back-end) rather than building from scratch.
- Explore memory + knowledge-graph flavored agent systems (and learn what breaks first).
- Run local/dev deployments (source or Docker) and iterate quickly.
Pricing snapshot (auto-enriched):
The code is free to use, but it’s licensed under AGPL-3.0 (important for commercial/network use). In practice, you’ll still pay for whatever LLM endpoint you configure (the project supports OpenAI-SDK-compatible APIs) and any optional memory services you wire in (the repo references Zep Cloud for memory/graphs). The docs also caution that runs can be compute/token heavy, so start small.
Work-use / compliance snapshot (auto-enriched):
License: AGPL-3.0 is a strong copyleft license, including “network interaction” obligations—if you modify and run it as a network service, you may need to provide corresponding source to users interacting with it. Plan for a legal review before you put this behind a login and call it “internal-only.”
Data handling: Your seed materials and prompts may be sent to third-party LLM providers and any configured memory backends. What gets stored, for how long, and where depends on your providers and your deployment configuration. Unknown / not confirmed.
Alternatives (auto-enriched):
- Microsoft AutoGen: a framework for building multi-agent apps and agent-to-agent collaboration. It’s great when you want orchestrated agent workflows, but it’s not specifically packaged as a “parallel world” simulation product.
- CAMEL-AI OASIS: a large-scale agent social interaction simulation framework (and MiroFish explicitly credits OASIS as an engine component). If you want to build your own simulation product/UX with fewer opinions baked in, using OASIS directly may be cleaner.
Before you adopt it:
- Do a tiny run first (short horizon / fewer rounds) and measure cost + latency before you promise anyone “real-time forecasting.”
- Decide where sensitive inputs go: pick an LLM provider + storage/memory layer that matches your org’s data policy.
- Confirm your AGPL posture: if you’ll host this for others (clients, colleagues, the public), get clarity on what you must disclose and how you’ll comply.
Sources
- https://github.com/666ghj/MiroFish
- https://github.com/666ghj/MiroFish/blob/main/README-ZH.md
- https://github.com/666ghj/MiroFish/blob/main/LICENSE
- https://github.com/microsoft/autogen
- https://github.com/camel-ai/oasis