Last updated: 2026-10-27
MiroFish is an open-source swarm simulation engine with 82,230 GitHub stars as of 27 October 2026. It turns a seed document such as a news report or policy draft into thousands of LLM-driven personas, lets them argue on two simulated social networks, then writes a prediction report you can question agent by agent.
About MiroFish
MiroFish is a swarm simulation engine published on GitHub in November 2025 by BaiFu, a developer whose profile lists USTC and BUPT, with strategic support and incubation from Shanda Group. It passed 80,000 stars inside a year and follows BaiFu's earlier BettaFish, a multi-agent public-opinion analysis assistant with 42,540 stars of its own. Reports that an investor committed around 30 million yuan to a company built on MiroFish have not been confirmed by the project.
You give it seed material (a report, a news story, even a novel) and a question in plain language. It builds a knowledge graph from the text, generates a cast of personas with their own memories, and runs them on two simulated social platforms at once. The simulation layer is CAMEL-AI's open-source OASIS, a social media simulator its authors scale to one million agents. When the run ends, a report agent writes up what happened, and you can chat with any persona to ask why it posted what it did.
Viral clips calling it a way to test thousands of fake customers on your site misread it. MiroFish never loads a web page or clicks a button. It models how people talk about something online, which suits crisis communications, policy announcements and campaign messaging. Teams that want to put their own AI agent through simulated conversations should look at Kashikoi instead. If you want agent teams that do real work, start with CrewAI or SuperAGI, and Eigent is a desktop multi-agent app built on the same CAMEL-AI base. Google Research's Vantage puts simulated AI personas to a much narrower use, scoring how one person collaborates with a team.
MiroFish is model-agnostic: any endpoint that speaks the OpenAI API format works. Its setup file points at qwen-plus, served by Alibaba Cloud, while endpoints from OpenAI and DeepSeek plug in the same way, and an optional second boost model can take the high-volume calls. The bill therefore grows with cast size and run length.
We file it with emerging AI software in the infrastructure and platform directory, next to other open-source tools, for a reason: treat its reports as rehearsals to think with, not forecasts to bet on.
Pricing
Free and open source, self-hosted only. Your running costs are the LLM API you connect (the default setup points at qwen-plus on Alibaba Cloud) plus Zep Cloud memory once you pass its free allowance. A hosted online edition has a waitlist and no published price yet.
Key Features
- Seed-to-world graph: Uploaded reports, news or fiction are parsed into a GraphRAG knowledge graph, and the entities and relationships in it become the starting cast.
- Personas with memory: Each generated agent carries its own persona plus individual and shared memory held in Zep Cloud, updated as the simulation clock advances.
- Two social platforms in parallel: Agents post, reply and react on a Twitter-style feed and a Reddit-style forum at the same time, driven by CAMEL-AI's OASIS engine.
- ReportAgent: After the run a dedicated agent with its own toolset queries the finished world and writes the prediction report.
- Interview any agent: You can open a chat with any individual persona in the simulated world, or question the ReportAgent about how it reached a conclusion.
- Bring your own model: Any OpenAI-format LLM endpoint works, with a separate optional slot for a faster model to absorb bulk calls.
Pros
- Every stage is open code, so you can read exactly how personas, memory and the report are produced, which closed synthetic-research services do not let you do.
- Short path to a first result: one Docker Compose command and two API keys start a run on a laptop.
- Being able to question individual agents after a run shows why the report concluded what it did, instead of handing you a single number.
Cons
- No published backtest: neither the README nor the website compares a simulation against a real outcome, so a report is a scenario rather than a forecast.
- When every persona runs on the same model, opinions can converge into a consensus a real crowd would not reach.
- Early software: the latest tagged release is v0.1.2 from 7 March 2026, and one maintainer wrote most of the commits.
- It simulates social media chatter, not your website, so it cannot click through a checkout or measure conversion.
- Spend climbs quickly with more agents and more rounds, and there is no built-in cost estimate before a run.
Data Handling
- Training-data policy
- Self-hosted. Seed material is sent to whichever LLM endpoint and Zep Cloud account you configure; the project runs no hosted service of its own yet.
Frequently Asked Questions
What does MiroFish actually cost?
The software itself costs nothing: there is no paid plan, and the hosted online edition is still a waitlist with no announced price as of 27 October 2026. Real spend comes from the LLM API you connect, since every persona calls the model in every round, plus Zep Cloud once you outgrow its free allowance. The project's own setup notes warn that consumption is high and suggest trying runs under 40 rounds first.
Can you use MiroFish without paying?
Yes. The code is free to run and modify under the AGPL-3.0 licence, which also means that if you offer a modified version to other people as a network service you must publish your changes. A public demo linked from the README replays a prepared simulation in the browser with no keys needed. Small experiments fit inside Zep's free quota, but the model calls are rarely free at a useful scale.
What are MiroFish's closest competitors?
OASIS from CAMEL-AI is the engine underneath, so researchers who want raw control without the interface can use it directly. CrewAI suits developers building agent teams that ship work rather than simulate opinion. Kashikoi is the better choice when the thing under test is your own AI agent, not a public audience.
How does MiroFish compare to CrewAI in 2026?
They solve different problems. CrewAI is an MIT-licensed framework you write code against to build agents that finish tasks. MiroFish is a ready-made app with one job, social simulation, under a stricter copyleft licence. If you need a rough read on public reaction, MiroFish gets you there in an afternoon; for anything else, CrewAI is the safer bet.
How do you set up MiroFish?
You need Node.js 18 or newer, Python 3.11 or 3.12 and the uv package manager. Copy .env.example to .env, add an LLM API key and a Zep Cloud key, run npm run setup:all and then npm run dev; the interface opens on port 3000 with the API on 5001. Docker users can run docker compose up -d instead. Keep the first simulation small so you see the cost before scaling up.
Top Alternatives
- CrewAI: Pick CrewAI if you need coded agents that complete real tasks; MiroFish only rehearses how a crowd might respond.
- SuperAGI: Choose SuperAGI when autonomous agents should run business workflows; MiroFish produces scenario reports, not finished work.
- Kashikoi: Pick Kashikoi to stress-test your own AI agent in simulated conversations; MiroFish simulates public opinion instead.
HokAI guides covering MiroFish
- What Is MiroFish? What the 82K-Star Swarm Simulator Really Does: MiroFish simulates AI personas arguing on fake social networks, not shoppers on your site. What a run costs at today's API prices, its bugs, and if it is legit.