MiroFish Setup-Guide: Starte deine erste autonome KI-Simulation

Set up MiroFish, the swarm intelligence engine with 79K GitHub stars, and run your first simulation. Repo link, step by step setup and 5 ready to paste prompts.

---

Warum dieser Ansatz für die Schweiz und den DACH-Raum entscheidend ist

Im Schweizer Hochlohnland entscheidet die Effizienz von Arbeitsabläufen über Wettbewerbsfähigkeit. Gleichzeitig verlangt das revidierte Schweizer Datenschutzgesetz (revDSG) strikte Kontrolle über vertrauliche Daten.

Statt für dutzende einzelne SaaS-Abos hunderte Franken pro Monat auszugeben, setzen führende Entwickler und smarte Betriebe 2026 auf Open-Source-Router, lokale Modelle und standardisierte Schnittstellen:

• 0 Fr. laufende Kosten: Maximale Nutzung der grosszügigen Free-Tiers und Open-Weights.
• 100% Datensouveränität: Sensible Keys und Abfragen bleiben verschlüsselt auf deiner eigenen Hardware.
• Kein Vendor-Lock-in: Fällt ein Anbieter aus, wechselt der Router in Millisekunden nahtlos zum nächsten Modell.

---

Schritt-für-Schritt Anleitung

The MiroFish Setup Guide

The repo link, a step by step setup and 5 prompts to run your first AI crowd simulation.

You commented SWARM, so here is the whole thing: the repo link, a step by step setup and 5 prompts you can paste straight into MiroFish. You do not need to be a developer to follow along. We go one step at a time.

What it is

MiroFish is a swarm intelligence engine on GitHub. You give it a "seed" (a news story, a report, a policy draft, even a novel) and describe what you want to know in plain words. It builds a digital world full of AI agents, each with its own personality and memory, lets them react and interact, and then writes you a prediction report. You can even chat with any agent in that world afterwards.

MiroFish, from the project's GitHub repo.

It was built by a computer science student in about 10 days, it is incubated by Shanda Group, and it has 79K stars on GitHub.

A simple way to think about it: instead of asking one AI "what will happen?", MiroFish asks a whole crowd of AI people and watches what they do. You see the reactions play out before the real world does.

Before you start

A Mac, Windows or Linux computer.Node.js 18 or newer from nodejs.org.Python 3.11 or 3.12 (not 3.13) from python.org.uv, the Python package manager, from docs.astral.sh/uv.An LLM API key from any provider that works with the OpenAI SDK format. The project recommends Alibaba's qwen-plus model.A free Zep Cloud API key from app.getzep.com. This is the agents' memory, and the free monthly quota is enough for simple use.

A Mac, Windows or Linux computer.

Node.js 18 or newer from nodejs.org.

Python 3.11 or 3.12 (not 3.13) from python.org.

uv, the Python package manager, from docs.astral.sh/uv.

An LLM API key from any provider that works with the OpenAI SDK format. The project recommends Alibaba's qwen-plus model.

A free Zep Cloud API key from app.getzep.com. This is the agents' memory, and the free monthly quota is enough for simple use.

Step 1: get the code

Open your terminal and run:

git clone https://github.com/666ghj/MiroFish.git cd MiroFish

git clone https://github.com/666ghj/MiroFish.git cd MiroFish

Step 2: add your keys

Copy the example settings file:

cp .env.example .env

cp .env.example .env

Open .env in any text editor and fill in these lines:

LLM_API_KEY=your_api_key LLM_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1 LLM_MODEL_NAME=qwen-plus ZEP_API_KEY=your_zep_api_key

LLM_API_KEY=your_api_key LLM_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1 LLM_MODEL_NAME=qwen-plus ZEP_API_KEY=your_zep_api_key

Using a different provider? Swap LLM_BASE_URL and LLM_MODEL_NAME for your provider's OpenAI-compatible address and model name.

LLM_BASE_URL

LLM_MODEL_NAME

The step people miss: the file also has three LLM_BOOST_... lines. If you are not using a second "boost" model, delete those three lines. The file itself says they must not be there unless you use them.

LLM_BOOST_...

Step 3: install and start it

From the MiroFish folder, run:

npm run setup:all npm run dev

npm run setup:all npm run dev

The first command installs everything in one go. The second starts the app. Open http://localhost:3000 in your browser and you are in.

Prefer Docker? After step 2, run docker compose up -d instead. It opens on the same address.

docker compose up -d

Step 4: run your first simulation

Upload a seed. A PDF, Markdown or text file: a news article, a report, a product announcement, a chapter of a story.Describe your question in plain words. What do you want to see happen? Use one of the prompts below to start.Let it build the world. MiroFish pulls the people and groups out of your seed, gives each agent a personality and memory, and runs the simulation.

Upload a seed. A PDF, Markdown or text file: a news article, a report, a product announcement, a chapter of a story.

Describe your question in plain words. What do you want to see happen? Use one of the prompts below to start.

Let it build the world. MiroFish pulls the people and groups out of your seed, gives each agent a personality and memory, and runs the simulation.

The start screen: upload your seed on the right, describe your question below it. Screenshot from the MiroFish repo (interface shown in Chinese).

Start small. The project warns that runs use a lot of AI credits, so try simulations with fewer than 40 rounds first.

The world it builds: a graph of everyone in your seed, and the agent personas it generated from them. Screenshot from the MiroFish repo.

Step 5: read the report and talk to the agents

When the simulation finishes, MiroFish's ReportAgent writes a full prediction report. Then the fun part: you can chat with the ReportAgent, or with any individual agent in the simulated world, and ask why they reacted the way they did.

A prediction report being written, section by section. Screenshot from the MiroFish repo.

Asking follow-up questions after the run. Screenshot from the MiroFish repo.

The kit: 5 prompts to paste

Paste one into the question box after you upload a matching seed. Swap the bits in brackets for your own.

1. Test a launch before you post it

This seed is our launch announcement for [product]. Simulate how [target audience] reacts on social media over the first week. Who gets excited, who pushes back, which objections spread fastest, and what would change their minds?

This seed is our launch announcement for [product]. Simulate how [target audience] reacts on social media over the first week. Who gets excited, who pushes back, which objections spread fastest, and what would change their minds?

2. Rehearse a crisis

This seed describes [incident]. Simulate how the news spreads online over 7 days, how public sentiment shifts, and how influential accounts react. Compare two responses: [response A] and [response B]. Which one calms things down faster?

This seed describes [incident]. Simulate how the news spreads online over 7 days, how public sentiment shifts, and how influential accounts react. Compare two responses: [response A] and [response B]. Which one calms things down faster?

3. Read the crowd on a news event

This seed is a news article about [event]. Simulate how [group 1], [group 2] and [group 3] react. Predict the main opinion trends and the biggest risk points for each group.

This seed is a news article about [event]. Simulate how [group 1], [group 2] and [group 3] react. Predict the main opinion trends and the biggest risk points for each group.

4. Pressure test a marketing idea

This seed is our campaign plan for [campaign]. Simulate how [audience] responds. Which message gets shared the most, which one gets ignored, and what topic takes over the conversation?

This seed is our campaign plan for [campaign]. Simulate how [audience] responds. Which message gets shared the most, which one gets ignored, and what topic takes over the conversation?

5. The fun one: finish a story

This seed is the first part of [story]. Simulate how the main characters act from here based on their personalities and relationships. Predict the most likely ending and two alternative endings.

This seed is the first part of [story]. Simulate how the main characters act from here based on their personalities and relationships. Predict the most likely ending and two alternative endings.

If something goes wrong

The app will not start? Check your Python version. MiroFish needs 3.11 or 3.12.Errors about the AI model? Open .env again: check your key, base URL and model name, and make sure the three LLM_BOOST_... lines are gone if you are not using them.Burning through credits? Cut the number of rounds and use a shorter seed.

The app will not start? Check your Python version. MiroFish needs 3.11 or 3.12.

Errors about the AI model? Open .env again: check your key, base URL and model name, and make sure the three LLM_BOOST_... lines are gone if you are not using them.

LLM_BOOST_...

Burning through credits? Cut the number of rounds and use a shorter seed.

The honest part (read this)

MiroFish is an early project (version 0.1.2), so expect rough edges. Its "predictions" are simulations of how people might react, not guarantees. The creator himself has said it is not meant to call the future exactly, but to turn possible futures into scenarios you can compare. Use it to explore and rehearse decisions, and never risk money on a simulation alone.

It runs on your own computer, but your seed and prompts are sent to your AI provider, so keep that in mind with private documents. The code is AGPL-3.0 licensed: free to use, but read the license before building a product on top of it.

Who this is for

Marketers, founders, creators and curious builders who want to see how people react before they hit publish. Set it up tonight and run your first world.

Repo: https://github.com/666ghj/MiroFish

Website: https://mirofish.ai

The AI Leverage is how a few thousand builders get sharper every week: the tools that quietly change the game, and the exact way to run them, so AI saves you real time and real money instead of just adding another tab. Come get better with us. by bestapps.ai on instagram.

Keep Reading

The AI Leverage Newsletter

---

Praxistipps für Schweizer KMU & Power-User

1. Verschlüsselte Schlüsselverwaltung: Speichere keine Produktiv-Keys in ungesicherten Umgebungsvariablen. Nutze lokale Secrets-Manager.
2. Fallback-Hierarchie: Konfiguriere schnelle Modelle (z.B. Groq / Llama 3) für Routine-Klassifizierungen und schwere Modelle nur für komplexe Synthesen.
3. DSG-Konformität: Stelle sicher, dass bei der Nutzung internationaler Endpunkte keine personenbezogenen Kundendaten im Prompt enthalten sind.

---

Fazit & Download

Dieser Workflow beweist: Professionelle KI-Infrastruktur muss nicht teuer sein. Mit den richtigen Open-Source-Bausteinen baust du in unter 15 Minuten ein Setup auf, das kommerziellen Agenten-Plattformen in nichts nachsteht.

Original-Inspiration von The AI Leverage: https://theaileverage.beehiiv.com/p/the-mirofish-setup-guide

Du willst mehr praxiserprobte KI-Strategien für dein Business?
Abonniere den kostenlosen ki-netic Newsletter auf ki-netic.ch oder vernetze dich mit uns auf Instagram @ki_netic_schweiz.