MiroFish is an open-source engine that turns a document into a simulated social network of AI agents and writes a forecast from what they do.
On 7 March 2026 the repository 666ghj/MiroFish reached number one on GitHub’s global trending list. According to Pedaily, a Chinese venture capital news site, Shanda Group founder Chen Tianqiao had already committed 30 million yuan to incubate it. By 12 October 2026 the project had 75,800 stars and 11,600 forks. It still has no published backtest of its forecasts against real outcomes.
Mind the domain. The most-promoted site, mirofish.my, is not run by the project. It is a paid service that describes itself as independent. The official code lives on GitHub, which lists mirofish.ai as the project’s homepage.
What is MiroFish and who built it?
MiroFish was written by Guo Hangjiang, known online as BaiFu, a final-year undergraduate at the time according to Pedaily. His earlier project, BettaFish, a multi-agent tool for analysing public opinion, also topped GitHub’s trending list in late 2025. Chinese tech outlet iFeng reports that he built MiroFish in ten days inside Shanda, where Chen had invited him after BettaFish took off. The README confirms Shanda’s incubation; the investment figure comes from the Chinese press.
The code is licensed under AGPL-3.0, a copyleft licence with specific obligations for anyone who runs a modified version as an online service.
How does MiroFish work?
You upload a “seed” document (a news story, a draft policy, an earnings note, even a novel) and ask a question in plain language. The README describes five stages:
- Graph building. Entities and relationships are extracted into a knowledge graph (GraphRAG).
- Environment setup. Each entity becomes a persona with a personality, background and starting opinions.
- Simulation. The agents post and react on two simulated platforms, one modelled on Twitter and one on Reddit, over several rounds, with memory updated each round.
- Report. A ReportAgent reads the simulation and writes the forecast.
- Interaction. You can chat with any agent or question the ReportAgent.
The social simulation itself comes from OASIS, an open-source simulator from the CAMEL-AI community, published on arXiv on 19 November 2024 by researchers from the Shanghai Artificial Intelligence Laboratory, Oxford, KAUST and others. OASIS scales to one million agents and includes a recommendation algorithm that decides which posts each agent sees. Its authors also found that their agents follow the majority opinion more readily than real people do.
One point matters more than the architecture: every agent is the same language model playing a different role. A thousand agents are not a thousand independent people.
To run it you need Node.js 18+ and Python 3.11 to 3.12 (or Docker), an LLM reachable through an OpenAI-compatible API, and a Zep Cloud account for agent memory. The README recommends Alibaba’s Qwen-plus and warns that consumption is high.
Does a swarm beat a single LLM call? The WPP test
WPP’s research lab ran the cleanest comparison published so far. The author, Nikos, a staff data scientist at WPP’s Satalia, asked MiroFish each morning before the US open whether the S&P 500 would close up or down and which five constituents would gain the most. A control arm sent the same seed and prompt to a single Gemini call, with no swarm, no memory and no tools.
Only one trading day produced a complete run: 14 May 2026.
| Question | MiroFish swarm (~25 agents) | Single Gemini call | Actual outcome |
| S&P 500 direction | Down | Down | Up 0.77%, from 7,444.25 to 7,501.24 |
| Top five gainers | NVDA, AMZN, META, GOOGL, MSFT | WMT, BABA, DE, AMAT, CAVA | CSCO +13.4%, JBHT +7.1%, APP +7.0%, TTWO +6.8%, F +6.7% |
| Correct picks | 0 of 5 | 0 of 5 (two not in the index) |
Source: WPP Research.
The swarm’s picks were essentially the index’s largest companies. Whenever both arms gave a direction, they agreed, which suggests the multi-agent layer contributed no signal the underlying model lacked. In the authors’ words, “the swarm added complexity without adding information.”
The operational findings are just as relevant for anyone evaluating the tool:
- the published Docker image was outdated and Chinese-only;
- a hard-coded token limit tuned for Qwen truncated Gemini’s output until the team switched to the lighter gemini-2.5-flash-lite;
- you cannot set the number of agents: the system derives it from the entities in your seed. The team wanted 100 and got about 25;
- Zep Cloud’s free tier ran out after three runs, killing a simulation mid-run;
- MiroFish does not browse the web, so a forecast is only as current as the seed you write.
The whole experiment cost about one dollar. One trading day is not a verdict, and the authors say so; they suggest scenario planning and qualitative synthesis as fairer tests.
This article is for information only and is not financial or investment advice.
Is mirofish.my the official MiroFish site?
No. mirofish.my reuses the project’s name, logo and demo videos, but its footer calls it an “independent simulation product” rebuilt in Next.js. Plans cost $9, $19 and $49 a month. Its terms of service, last updated on 15 March 2026, name no operating company or jurisdiction. The homepage advertises “1M+” parallel agents per run; the original project claims thousands, the million is OASIS’s ceiling, and WPP got 25.
We found no source-code link on the site. Under AGPL-3.0, a modified version offered over a network must make its source available to users; a ground-up rewrite carries no such duty. From the outside it is impossible to tell which applies here, and we make no claim either way.
| Option | What you pay | Where your documents go | Operator |
| Self-hosted official repo | LLM tokens + Zep Cloud beyond the free tier | Your chosen LLM provider and Zep Cloud | You |
| Official online demo | Nothing | No upload: a pre-built simulation | MiroFish project |
| mirofish.my | $9 to $49 per month | Not disclosed | Not disclosed |
Sources: GitHub README; mirofish.my pricing page, checked 12 October 2026.
Where does your data go?
The README’s example configuration points to dashscope.aliyuncs.com, which Alibaba Cloud documents as its Beijing region, with data stored in mainland China. For organisations under the GDPR this matters: the EU has no adequacy decision for China, so personal data in a seed document would fall under the GDPR’s international transfer rules. Similar questions arise under other data-protection regimes. The fix is a configuration change, since any OpenAI-compatible endpoint works, including Alibaba’s Singapore region or a local model.
Zep Cloud, however, cannot be swapped out. WPP describes it as a hard dependency, so it needs the same vendor review as any cloud processor.
What is MiroFish actually good for?
The useful output is the map, not the number: a graph of actors and tensions, and personas you can question. That makes it a structured brainstorming aid for crisis communications, policy reaction or launch messaging, with findings checked by conventional research.
Keep three limits in mind. Agent count is not a sample size, because every agent comes from the same model and the same seed. Agents tend to herd, as OASIS’s own authors measured. And the simulated world knows nothing beyond the document you upload.
FAQ
Is MiroFish free? The software is, under AGPL-3.0. Running it requires a paid or local LLM and a Zep Cloud account with a limited free tier.
What is the official MiroFish website? The project lives at github.com/666ghj/MiroFish, which lists mirofish.ai as its homepage. mirofish.my is an independent paid service.
Can MiroFish predict the stock market? There is no evidence it can. In the only documented independent test, on 14 May 2026, it called the S&P 500 direction wrong and picked none of the day’s top five gainers.
How many agents does MiroFish use? The system decides, based on how many entities it extracts from your document. In WPP’s test that was about 25.
Where are uploaded documents processed? With the README’s suggested setup, by Alibaba Cloud in Beijing and by Zep Cloud. The LLM provider can be changed; Zep Cloud cannot.
Sources
- 666ghj/MiroFish README and licence, GitHub, accessed 12 October 2026
- 陈天桥投的AI产品登顶Github, Pedaily (via TMTPost), March 2026
- 中国最年轻首富,又杀回来了, iFeng Tech, 2026
- MiroFish: Is swarm intelligence worth the cloud bill?, WPP Research, 2026
- OASIS: Open Agent Social Interaction Simulations with One Million Agents, arXiv, 19 November 2024
- OASIS project page, CAMEL-AI
- Batch interfaces compatible with OpenAI, Alibaba Cloud Model Studio documentation
- Pricing and Terms of Service, mirofish.my, accessed 12 October 2026
