You spend three months building. Launch. Crickets. Or worse: feedback revealing that your pricing is wrong, your positioning confuses people, or they simply don’t care about the problem you solved.

This kill cycle takes down more solo builders than lack of technical skill ever will.

The old alternative was qualitative research—interviews, surveys, focus groups. They work, but they’re time-intensive and plagued by social desirability bias: people tell you what they think you want to hear.

MiroFish Offline offers something different. Instead of asking real people, you run a simulation with thousands of AI agents—each with distinct personalities, memory, and independent behavior—and observe how they react to your idea. The output isn’t a linear prediction. It’s an emergent pattern: consensus, controversy, specific objections, and priorities.

This guide walks you through running MiroFish locally (no coding required), creating a real-world validation simulation via visual interface, interpreting results, and using them to decide what to build—before you invest any real resources.

What You’ll Get

By the end of this guide, you’ll have:

  • MiroFish running locally on your machine (via Docker Compose)
  • Access to an intuitive web interface (zero lines of code)
  • A repeatable workflow for testing any business hypothesis before execution
  • Complete privacy—everything runs offline, nothing leaves your machine

Setup time: 5 minutes (Docker does the heavy lifting)
Operating cost: Zero
Required profile: Have Docker installed, know how to click a browser button

Why MiroFish Offline Instead of the Cloud Service

Before diving into setup, it helps to understand what MiroFish is and why the offline version matters.

What Is MiroFish

MiroFish is a swarm intelligence engine originally created by researcher “Bifu” with support from the Shanda Group. The core idea: simulate a digital society with thousands of autonomous agents, each with personality and memory, to observe how collective behaviors emerge from an initial stimulus—what they call a “seed.”

Under the hood, the system uses GraphRAG (Graph Retrieval-Augmented Generation) to model complex relationships between agents, contexts, and behaviors. Each agent doesn’t simply respond to a prompt—it responds to its memory history, interactions with other agents, and the current environmental state.

How This Differs from Asking ChatGPT

Ask ChatGPT “will people like my product?” and you get a linear, optimistic answer with no friction. The model tries to please you.

Run your idea through MiroFish and you get something else entirely: 500 agents with distinct personalities processing the same idea independently. Some get excited. Others ignore it. Others raise specific objections. The distribution of these responses, the patterns of agreement and disagreement, the topics that emerge unbidden—that’s the real signal.

The difference isn’t philosophical. It’s practical:

ApproachWhat It Produces
Simple ChatGPT queryOne perspective, optimized to please
Survey with real peopleSocial desirability bias, skewed answers
MiroFish OfflineDistribution of emergent behaviors

Why the Offline Version

The online MiroFish service has a cost and requires sending your business hypotheses to external servers. The offline version—the nikmcfly/MiroFish-Offline repository on GitHub—runs local models via Ollama and stores the knowledge graph in local Neo4j. You gain:

  • Zero cost per simulation
  • Total privacy of your ideas
  • Control over which models run
  • Ability to run simulations in parallel without throttling

The trade-off? Docker needs to be installed. That’s it.

Minimum Requirements

  • Docker Desktop installed (Windows, Mac, Linux)
  • 16 GB of RAM (8 GB works, but slower)
  • 50 GB of disk space (for AI models)
  • Internet connection (to download Docker images and models initially)
  • macOS, Windows, or Linux (Docker runs on all)

If you already have Docker, you’re ready to go.

Setup in 5 Minutes

Step 1 — Clone the Repository

git clone https://github.com/nikmcfly/MiroFish-Offline.git
cd MiroFish-Offline

Step 2 — Copy Default Configuration

cp .env.example .env

This creates the .env file with correct variables. You don’t need to edit anything—the defaults work perfectly.

Step 3 — Start Everything with Docker Compose

docker compose up -d

This starts:

  • Flask Backend (MiroFish API) on port 5001
  • Vue Frontend (web interface) on port 3000
  • Neo4j (graph database) on port 7474
  • Ollama (local AI models) on port 11434

The process takes 2 to 5 minutes depending on your connection.

Step 4 — Download AI Models

While Docker initializes, pull the models in another terminal:

docker exec mirofish-ollama ollama pull qwen2.5:32b
docker exec mirofish-ollama ollama pull nomic-embed-text

This downloads:

  • qwen2.5:32b (~18 GB) — the primary language model
  • nomic-embed-text (~250 MB) — embedding model

Download takes 10 to 30 minutes depending on your connection. You can let it run in the background.

Step 5 — Open in Browser

http://localhost:3000

Done. The interface is live.

Understanding the Interface

MiroFish operates through 5 stages in the web UI:

1. Document Upload

On the home page, click “New Project” and upload a file:

  • PDF
  • Markdown
  • Text file

Or describe your idea directly in the text field.

2. Generate Ontology

MiroFish analyzes your document with AI and extracts:

  • Key entities (product, market, customer)
  • Relationships between them
  • Context

This creates a “schema” that agents will use to understand the simulation world.

3. Build Knowledge Graph

The system constructs a Neo4j graph with extracted entities. This gives agents “memory” of interactions and awareness of each other.

4. Generate Agent Profiles

MiroFish automatically creates agent profiles. Each one has:

  • Distinct personality
  • Values and priorities
  • Behavioral history

You can configure 50 to 500 agents. More agents = more perspectives = more signal (but slower execution).

5. Run Simulation

Click “Start Simulation”. The system:

  • Activates agents
  • Allows them to interact with each other
  • Simulates social network behavior (Twitter, Reddit-style)
  • Generates automated report

The result is comprehensive analysis showing:

  • Overall sentiment
  • Primary objections
  • Agent profiles most likely to adopt/reject
  • Emergent patterns

Practical Example: Validating an AI Thumbnail SaaS

Let’s work through a concrete example: “SaaS that generates AI-optimized YouTube thumbnails for creators.”

What You Do:

In the web interface:

  1. Click “New Project”
  2. In the description field, write:
Product: ThumbAI
Description: SaaS that analyzes a YouTube channel's performance history, 
identifies thumbnail patterns with highest CTR, and generates new 
AI-optimized thumbnails using generative AI.

Price: $29/month

Target Audience: Creators with 1k to 100k subscribers

Value Proposition: Increase your CTR by 30% without needing a designer 
or understanding image editing.

Context: 2026, the AI tools for creators market is saturated, but most 
focus on text content. YouTube has 50M+ creators globally. Mid-tier 
creators (10k-100k subscribers) feel performance pressure but lack budget 
to hire designers.
  1. Click “Generate Ontology”
  2. Once complete, click “Build Knowledge Graph”
  3. When ready, click “Create Simulation”
  4. Configure agents (or use the defaults)
  5. Click “Start Simulation”

What Happens:

MiroFish simulates:

  • 200 YouTube creators of varying sizes
  • 50 content agencies
  • Mix of skeptics and early-adopters

They process the ThumbAI idea and interact over 10 rounds.

Reading the Results

After the simulation (15-30 minutes), you receive an automated report with:

Overall Sentiment

Positive:   58%
Neutral:    28%
Negative:   14%

What it means: Genuine interest exists, but real objections are present.

Primary Objections

1. "How do I know the AI understands my channel's style?" (67% of agents)
2. "$29/month is expensive if I don't see results in 30 days" (54%)
3. "I've tried other thumbnail tools that promised this" (48%)
4. "I want to see real examples from channels like mine" (43%)

What to do: These are the objections you address in landing page copy, onboarding flow, and value positioning.

Segments with Highest Adoption

Growth-Obsessed Creators (20k-100k subscribers): 82% positive
Casual Creators (<10k): 35% positive
Content Agencies: 76% positive

What it means: Start with growth-focused creators. Save casual creators for later.

Price Sensitivity

$19/month:    74% acceptance
$29/month:    58% acceptance
$49/month:    31% acceptance
Freemium:     68% preference

What to do: $29 is solid, but a freemium model (with generation limits) might lower adoption friction.

Features Agents Requested

1. Free trial period (91%)
2. Dashboard showing before/after CTR comparison (78%)
3. Real examples of channels using the tool (72%)
4. YouTube Studio integration (65%)
5. Templates by content niche (57%)

What to do: These features go on your roadmap. #3 is critical—social proof drives adoption.

How to Interpret Without Fooling Yourself

✅ Use Results For:

  • Deciding if it’s worth building — 58% positive with clear objections = worth exploring, but don’t build without addressing them
  • Refining positioning — If 67% ask “how does AI understand my style?”, that’s your core selling point
  • Prioritizing features — Features mentioned by 70%+ of agents are MVP-critical
  • Choosing initial segment — Growth-focused creators (82%) before casual creators (35%)
  • Testing price — The drop from 58% to 31% between $29 and $49 signals your ceiling

❌ Don’t Use For:

  • Guaranteeing success — Simulation signals risk, not proof of market fit
  • Ignoring objections — If 67% raise the same concern, you must address it
  • Taking absolute numbers literally — 58% positive doesn’t mean “58 of every 100 will buy”
  • Skipping real conversations — Always validate with 5-10 real people from your target segment before launch

Practical Applications for Solo Builders

1. Validate Ideas Before Coding

The most direct use case. Simulate before you write a line of code. 30 minutes beats 3 months of wasted development.

Example: A builder planned 3 months for a report automation SaaS. Simulated first. Discovered the core barrier wasn’t the product—it was the perception that “another BI tool” offers no differentiation. He shifted positioning, developed the MVP in 6 weeks with messaging already tested.

2. Test Messaging Variants Pre-Launch

Create different product descriptions with different angles:

  • Variant A: “Report automation for agencies”
  • Variant B: “Client-ready reports, zero manual work”

Simulate both. Compare objections and sentiment. The winner goes straight to your landing page.

3. Surface Objections Before Building Onboarding

The simulation’s primary objections become your onboarding roadmap. If 67% ask “how does AI understand my style?”, your first onboarding step answers exactly that.

4. Calibrate Your Pricing

Test 3-4 price points and watch acceptance drop. Where it falls >20% is your ceiling.

5. Test Pivots Without Derailing Momentum

Considering a pivot? Simulate both scenarios and compare. Saves you from pivots that kill momentum.

6. Preview Content Reception

Creating content for growth/acquisition? Simulate audience reaction before investing weeks.

7. Uncover Non-Obvious Use Cases

Dig into detailed agent logs. Often they use your imagined product in ways you didn’t anticipate. A “content agency rep” agent might reveal a multi-seat plan opportunity you missed.

Monetization: What You Can Build With This

Model 1 — Validation as a Service

Offer validation services to other builders and founders. You run the simulation, interpret results, and deliver a 5-10 page report.

Price: $500 to $1,500 per validation
Volume: 4-8 validations/month with analysis
Revenue: $4,000 to $12,000/month
Advantage: Zero operational cost, just your time

Model 2 — Validation SaaS Product

Build a polished interface on top of MiroFish and sell access to startups and builders.

Price: $49 to $99 per simulation, or monthly subscription
Model: Monthly quota of 5-10 simulations
Variable Cost: Zero (runs locally)
Revenue: $2,000 to $5,000/month with 10-20 customers

Model 3 — Product Strategy Consulting

Combine MiroFish simulations with deep product/market knowledge. You don’t just simulate—you recommend strategic adjustments and roadmap prioritization.

Price: $2,000 to $5,000 per project
Duration: 2-4 weeks per engagement
Ideal Customer: Technical builders going from zero to market

Limitations and When NOT to Use

Model Bias

Agents are generated by the LLM (qwen2.5:32b by default). Their behaviors reflect patterns in the model’s training data—which carries geographic, cultural, and temporal bias.

Behaviors in highly specialized, very new, or culturally distinct markets aren’t modeled with full fidelity.

Mitigation: Run simulations with multiple models and compare. If qwen2.5 and mistral reach the same conclusions, the signal is stronger.

Simulation Can’t Predict Black Swans

Unexpected events (algorithm changes, regulatory shifts, dominant competitor entering) don’t appear in simulation.

The simulation models behaviors based on the context you provide. Incomplete context = incomplete results.

Garbage In, Garbage Out

A vague seed produces vague results. If you describe your product as “marketing SaaS for businesses,” agents process a fuzzy idea.

Spend time on your product description. The more specific (market, audience, problem, solution), the better the signal.

Doesn’t Replace Real Conversations

Simulation reduces risk; it doesn’t eliminate uncertainty. Before launching anything with meaningful price, validate with 5-10 real people from your simulated target segment.

The simulation shows where to look. Real conversations confirm what you found.

1. Raw idea
        |
        v
2. Initial simulation (5-10 min)
        |
        v
3. Review objections and adoption clusters
        |
   [Does the idea signal positive?]
   No  → Refine positioning or kill it
   Yes → Continue
        |
        v
4. 5-10 real conversations with people in simulated segment
        |
   [Does signal validate?]
   No  → Return to step 2 with new hypothesis
   Yes → Build the MVP
        |
        v
5. MVP with messaging already calibrated by simulation

This complete cycle—from raw idea to validated MVP—can run in 1-2 weeks for digital products.

Next Steps

You have MiroFish running. The immediate next step is to take a real decision you need to make—a product you’re thinking of building, a pricing question you need to solve, a positioning shift you’re considering—and run an actual simulation.

The quality of your result depends on the quality of your product description. Spend 15-30 minutes writing a clear description of your product, target market, ideal customer, and value proposition.

For deeper ideas on idea validation using complementary approaches, see How to Validate Your SaaS Idea Before Coding—it covers qualitative research techniques that work alongside simulation. And if you’re still deciding which ideas are worth simulating, How to Validate a Digital Product Idea has the pre-screening framework.

MiroFish Offline gives you the laboratory. What you put into it determines the value you extract.