What is MiroFish? Using AI to simulate customer feedback before building a product

If you are considering an app that combines orders, MiroFish can help explore what might concern different shops. The OrderBoard example shows how to frame a simulation and turn its findings into questions and evidence to gather from real shops.

Illustration of people in different roles gathered around a briefing sheet.

What MiroFish simulates

MiroFish is an open-source project that lets AI agents interact based on source material. Agents take different roles, such as shop owner or administrator, so they may raise different issues about the same proposal. Product teams can use the results to build a list of concerns and research questions.

The agents come from the source material and a language model, so they cannot serve as a sample of customers in the market. If ten agents agree on a price, that count describes what happened in that simulation run.

Here, an agent is software that uses AI to play a role and decide how to respond or act. Imagine a small shop owner, a manager with several branches, and a bookkeeper all hearing about a new app. Each may notice something different and respond to the others. That is the kind of interaction a multi-agent simulation models.

A single chatbot answers on its own. MiroFish gives agents roles, relationships, and memory to draw on over several rounds. The project describes a setup that uses language models, Zep for graph and memory services, and OASIS for simulation. Read the project's “Predict Anything” tagline in that context. A simulation cannot estimate how many customers hold an opinion or establish that a product will succeed.

How MiroFish turns documents into a simulation report

  1. Build a knowledge graph. The system extracts people, organizations, and relationships from the documents to give agents a common starting point.
  2. Set up the environment. The system assigns roles and gives each agent context for its responses.
  3. Simulate interactions. Agents post or respond in rounds in an environment powered by OASIS. Results can vary between runs.
  4. Generate a report. ReportAgent summarizes what happened in the simulation and which issues surfaced.
  5. Ask follow-up questions. Talk to an agent or ReportAgent to understand where an observation came from.

A round may include several interactions, so the round count differs from the number of posts or agents. The report draws on the source documents, knowledge graph, and events from the run. Check any claim about the real world against an outside source.

Describe the product and how shops work today

Start with a short document. Say where orders arrive, who copies the details, and where mistakes tend to happen. Distinguish one-person shops, shops with staff, and shops with several locations. Describe their existing tools, such as notebooks or spreadsheets, and what those tools do well, so the scenario accounts for reasons a shop might stay with them.

Use concrete conditions, such as “It combines orders from two channels and has no price yet.” A claim like “most shops like it” makes it hard to tell whether the report found something new or repeated the input.

For this example, distinguish one-person shops, shops with three to five staff, and shops with several branches. Explain that orders can go missing when they come in through multiple chat channels. The proposed app combines orders and flags unfinished work, with initial setup and support for only some platforms.

The FAQ lists .pdf, .md, .markdown, and .txt as backend formats. The upload picker may list fewer. For a scanned file, extract the text and check it before uploading.

The OrderBoard example, from objections to evidence

OrderBoard is a fictional order-management app. Use the scenario to frame a simulation and prepare tests with real shops. The objections in the table are KrobClick's hypotheses.

Suppose OrderBoard can combine orders from some channels and flag unfinished work. Shops need to set it up, and there is no price yet. Explore what might concern a solo shop owner, a shop with staff, and a shop with several locations. For each concern, say whether it came from the source material or arose in the simulation, then suggest one question to ask a real customer.

Hypothetical objectionQuestion for a real shopEvidence to retain
Setup sounds like more work than our current process.Where did you get stuck the last time you adopted a new tool?Give the shop a sample order and record setup time and where help is needed.
We don't know who can access customer data.Who should be able to view customer details, change prices, or issue refunds?Interview the owner and administrator separately, then compare their answers with the prototype's policy and permissions.
The subscription may not pay for itself.Which repeated tasks currently take time or cost money?Time the same task with and without the prototype, then test a proposed price.
Our spreadsheet already does enough.What does the spreadsheet do well enough that you want to keep it?Have the shop complete the same task both ways. Compare time and errors.
We still have to copy orders if it misses our main channel.Which channels did yesterday's orders come through, and did any go missing?Trace permitted sample orders and prioritize channels based on the observed workflow.

Before scheduling interviews, check whether each concern came from the input or arose in the simulation, who raised it, and whether it recurs when you change one assumption at a time. For example, make old records easier to migrate. Take the concerns most likely to change your decision into interviews or prototype tests.

When to use MiroFish or collect evidence from people

MethodWhat it can answerWhat the answer covers
Single chatbotBrainstorms initial questions or concerns.Suggestions from a language model.
MiroFishExplores how different roles might respond to one another in a hypothetical situation.Results apply to the simulated conditions.
InterviewExplains a shop's reasons, experience, and constraints.Depends on who was interviewed and each shop's circumstances.
SurveyMeasures response patterns in a real sample.Quality depends on sampling and questionnaire design.
Prototype testShows where people get stuck and what help they need.A small test reveals friction. Broader adoption takes more evidence.

Try the demo or install MiroFish to run your own scenario

Start with the demo

The project's demo uses a prepared scenario. Look at its reports and conversations before setting up API access. To submit a new scenario, follow the Quick Start for your version.

Open the MiroFish demo

Install from source

The Quick Start lists Node.js 18 or later, Python 3.11 or 3.12, and uv. The example below uses a Bash-style shell. Start by cloning the project and copying its configuration template.

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

Open the .env file and enter the key, API endpoint, and model name for a provider compatible with the OpenAI SDK, along with your Zep Cloud key. Replace the placeholders below with your own values.

LLM_API_KEY=YOUR_MODEL_API_KEY
LLM_BASE_URL=YOUR_PROVIDER_API_BASE_URL
LLM_MODEL_NAME=YOUR_SUPPORTED_MODEL
ZEP_API_KEY=YOUR_ZEP_API_KEY

From the project directory, install the dependencies and start the web interface and backend.

npm run setup:all
npm run dev

The README puts the web interface at http://localhost:3000 and the API at http://localhost:5001. If you use Docker Compose, prepare the same .env file and start the services with the command below. KrobClick checked these commands against the project documentation. We have yet to run the installation ourselves.

docker compose up -d

Costs, data handling, and the limits of a simulation

The source is available under AGPL-3.0. Model calls and Zep Cloud quotas depend on your providers and usage. Start with a short document and a small number of agents and rounds, then compare usage before and after the run, including report generation. A ChatGPT subscription or Codex sign-in cannot substitute for the model API key in this setup.

The web interface runs on your machine, but the documented setup calls external APIs. Use fictional data for the first run, check what each service receives and retains, and keep the .env file out of Git and public screenshots.

An objection may recur because agents share the same source material or model. Use recurring concerns to prioritize what to test. Compare even a polished report with what customers actually say or do before making a decision.

Leave out customer names, phone numbers, payment details, private conversations, and confidential material you lack permission to share. Agent count, rounds, document length, and report generation all affect API usage. Check actual usage with your provider before estimating a budget per run.

If you find a similarly named site, follow the project's GitHub links to mirofish.ai or the demo listed in the README and FAQ.

Who should try MiroFish, and where others can start

ReaderHow to begin
Students and AI learnersUse the demo to see the workflow and reports before paying for API access.
Product teams looking for hypothesesExplore objections, then pick which ones to test with customers.
Developers familiar with Python, Node.js, and APIsRun from source with fictional data. Watch usage quotas and errors.
Teams trying to find out whether customers will buyStart with interviews, a prototype test, or real usage data.
Organizations with sensitive dataCheck where the data goes and each service's terms before sending it.

KrobClick would judge a run by how many specific questions it gives the team to test. Even when a report sounds confident, you still need to ask customers what they think.

Questions before trying MiroFish

Is MiroFish AI or customer survey software?

It is a simulation built on language models. Agent responses can help form hypotheses. Survey results come from real respondents.

Is MiroFish free to use?

The source is available under AGPL-3.0. Model and Zep costs depend on quotas, service terms, and usage.

Can I use MiroFish without programming experience?

You can open the demo and read the prepared reports. Running your own scenario with the documented setup takes a terminal installation and API configuration.

Can it tell me whether customers will like a product?

Use the simulation to prepare questions. Test them through interviews, surveys, or a prototype with real customers before drawing a conclusion.

How can I judge the report's credibility?

Check the source material, separate facts from model output, and rerun with a changed assumption. Confirm the points that matter through interviews or outside evidence.

Choose a first question you can simulate and verify

If you are planning a product, pick one problem and describe what you know in neutral terms. Take three to five objections that could change the decision to the people doing the work. Keep what they say and do separate from the simulation notes so you can see what changed the team's mind.

Sources

This article draws on the project's documentation. KrobClick uses a fictional shop scenario to illustrate how to frame questions.