What Is MiroFish? What the 82K-Star Swarm Simulator Really Does

MiroFish had 82,230 GitHub stars on 27 October 2026, eleven months after its first commit, and the most-shared description of it is wrong.

Short videos pitch it as a way to point thousands of fake customers at your website and watch them shop. The code does something else: MiroFish turns a document into a cast of AI personas and has them argue on two imitation social networks. If you run marketing, communications or a small product team, that difference decides whether this is a cheap rehearsal tool or a lost afternoon.

This guide reads the source, prices a run at today's API rates, and names what to use if you really do want synthetic visitors on a website.

What MiroFish actually does with your document

You give it two things: a seed document and a question written in plain language. The project README describes five stages, and the backend code matches them:

1. Graph building. The seed is cut into 500-character chunks and stored in Zep Cloud, a hosted memory service.
2. Environment setup. Each entity found (a person, a company, a ministry) becomes one agent.
3. Simulation. The agents act on a Twitter-style feed and a Reddit-style forum at once.
4. Report. A ReportAgent queries the finished world and writes the prediction.
5. Interaction. You can chat with any agent, or question the ReportAgent.

The social engine underneath is not MiroFish's own. It is OASIS from CAMEL-AI, which its authors describe as a social media simulator built for up to one million agents. MiroFish's contribution is the pipeline wrapped around it: seed in, personas out, report at the end. The same CAMEL-AI base also powers Eigent, a desktop app that points multiple agents at ordinary office work instead.

One detail shapes everything below. The agents are not sampled from a population. They are the people and organisations named in your document, so a press release about your product produces a cast drawn from that press release.

What changed between the viral claim and the code

Nothing in MiroFish opens a browser. The simulation script gives Twitter-style agents six possible actions (post, like, repost, quote, follow, or do nothing) and Reddit-style agents a dozen, such as commenting, voting, searching and following. There is no page load, no click, no form and no checkout anywhere in the backend.

The project never claims otherwise. Not once. The use-case list on mirofish.ai runs to six entries: financial sentiment, policy reaction, crisis PR, marketing message testing, fiction, and academic research. Every one of them is about how people talk, not how they behave on a page.

So where did the claim come from? The closest real use is the marketing entry: seed a campaign brief and watch simulated users discuss it. That is message testing in a fake comment section. It is not conversion testing, and it cannot tell you why visitors leave your pricing page.

If you want fake customers on your site, use these instead

The job the viral clip describes exists, and it belongs to a different category of software. Browser agents drive a real web page; testing platforms check flows; surveys ask real people. The table maps each job to a tool listed on HokAI.

Browser Use is the closest thing to what the clip promised: describe a task in one sentence and an agent tries to complete it on your live site. Where it fails is useful evidence about your interface.

Skyvern covers the same ground if you want to host it yourself, and Browserbase supplies the browsers when you want dozens running in parallel. None of them simulates a buyer's motives, which is the honest limit of the whole idea.

For flows rather than personas, the AI software testing category holds tools such as Kusho, which generates API and UI test suites, and QualGent for mobile builds.

And when the question is "would people buy this", the cheapest honest answer is still a short survey of real people through something like Typeform. Five real replies beat five hundred generated ones on the one thing that matters: they come from someone who might pay you.

How many model calls one run makes

MiroFish's cost is set by a handful of defaults in its config generator, and none of the coverage we read works them out. Here is the arithmetic from the source.

A default run simulates 72 hours at 60 minutes per round, so 72 rounds per platform. That is already above the README's own advice to start below 40 rounds. Each round, the simulation runner wakes a random slice of agents, and every agent it wakes makes one model call.

How big is the slice? The fallback settings in the config generator wake between N/15 and N/5 agents per round, where N is the cast size. Evening rounds (19:00 to 22:00) multiply that by 1.5, and the small hours (00:00 to 05:00) by 0.05.

• Average wake rate: about 0.13 × N agents per round.
• Weighted across a simulated day: about 2.7 × N calls per platform.
• Over 3 simulated days on 2 platforms: about 16 × N calls.

A 100-agent cast therefore lands near 1,600 simulation calls, before persona generation, the config step and the report agent's own tool calls and reflection rounds are added on top. That is not much. A 1,000-agent cast lands near 16,000. Per-agent activity levels and active hours trim those numbers, so read them as an upper estimate rather than a measurement.

One oddity turned up along the way. The config also defines morning and working-hours multipliers (0.4 and 0.7), but the runner only reads the peak and overnight values, so those two settings currently do nothing.

What a MiroFish run costs in dollars

Tokens per call depend on the persona length and how busy the feed is, and MiroFish does not log a standard figure. So the table below makes one stated assumption: 3,000 input tokens and 250 output tokens per agent turn, for the 1,600 turns of a 100-agent default run. Prices are OpenRouter's listed rates on 27 October 2026.

Double the tokens per turn and every figure doubles; multiply the cast by ten and so does the bill. On a small model, though, a default run costs a few dollars, not hundreds. Two of the four rows come from Google and OpenAI, and both sell larger models that would multiply the figure. Our cheapest LLM API guide covers what each of those low prices leaves out, and the model leaderboard shows how the same models compare on quality.

The README recommends Qwen-Plus through Alibaba Cloud's Bailian console. Alibaba's own pricing page sat behind a human-verification wall when we tried it, so the Qwen row above uses OpenRouter's rate, not Alibaba's direct one. Our explainer on what Qwen is actually for covers the model family.

The bill most people miss is memory, not tokens

The second cost is Zep Cloud. Zep's pricing page gives the free plan 10,000 credits a month, and a credit buys one stored "episode" of up to 350 bytes. MiroFish writes your seed into Zep in 500-character chunks, then writes agent actions back in batches of five.

For an English seed, a 500-character chunk is about 2 credits. Chinese text takes 3 bytes per character in UTF-8, so the same chunk costs about 5. On our arithmetic, a 100-agent default run uses several hundred to roughly a thousand credits, which puts the free plan at a handful of runs a month. The paid Flex plan is $125 a month for 50,000 credits.

Pasquale Pillitteri's analysis of MiroFish reports a WPP Research test in which the free Zep tier ran out after three runs. We could not open WPP's original post to confirm the detail, but it fits the arithmetic.

Is MiroFish legit?

Yes, as software. The repository is real, active and readable, and the code does what the README says. The more useful question is whether its output means anything, and that answer is less flattering.

What checks out:

• Open source. The licence is AGPL-3.0: free to run and modify, but a modified version offered as a network service must publish its changes.
• A real maintainer. BaiFu wrote 266 of the roughly 300 commits. His earlier project, BettaFish, a public-opinion assistant, has 42,540 stars.
• Named backing. The README credits strategic support and incubation from Shanda Group.
• Steady releases. Tagged versions arrived on 22 December 2025, 22 January 2026 and 7 March 2026.

What does not check out yet:

• No accuracy record. Neither the README nor the website compares a single simulation with a real outcome.
• Funding. Reports of roughly 30 million yuan from Shanda founder Chen Tianqiao circulate widely, and YourStory notes they have not been officially confirmed. Treat the figure as reported, not established.
• Version 0.1.2. Nothing has been tagged since March, and the newest commits on the main branch are automated star-count updates.

Who it affects: the people MiroFish helps and hurts

It helps people whose question is about conversation. A communications lead planning a sensitive announcement can see which objections surface and which groups pick a fight. A social-science researcher gets a configurable sandbox for studying how information spreads, which is what OASIS was built for.

It also helps anyone curious about multi-agent systems, because the whole pipeline is readable Python. If you plan to build agents that do real work rather than simulate opinion, frameworks like CrewAI or SuperAGI are the better starting point. The wider set sits in the AI orchestration category, and our roundup of AI agent builders compares the hosted options.

It hurts the people who read its report as a forecast. Three problems stack up:

1. One model plays every role. A thousand agents running on one model share its blind spots, so agreement can be an echo rather than a signal.
2. Results move between runs. Issue 17, opened on 6 January 2026 and still open, reports that the same material with a different model can reach a different conclusion.
3. The report sounds certain. The output reads like an analyst's brief whether or not the simulation supports it.

It also risks hurting the vendors that build synthetic research panels on real survey data. Their pitch is calibration; MiroFish makes the uncalibrated version free. Buyers who cannot tell the two apart will judge both by the cheaper one.

The known bugs, and the one that was quietly fixed

The sharpest criticism in current coverage is issue 529, opened on 12 April 2026. It shows the ReportAgent writing its own fake <tool_result> blocks, so a report could quote usernames and statistics that never existed in the simulation. The issue is still marked open, and the analyses we read describe the flaw as live.

The code says otherwise. It was patched. A commit on 22 July 2026 titled "fix: sanitize fabricated tool results everywhere" added a filter that strips model-written tool results before they reach the report, with tests covering nested, unclosed and upper-case variants. If you cloned before late July, pull again.

The fix narrows the problem; it does not close it. A model can still misread real tool output, so quotes in a report should be checked against the agent logs before anyone repeats them.

Validation is the bigger gap. Pull request 813, "Historical backtesting framework", would let a run's prediction be recorded and later scored against what really happened. It was opened on 15 September 2026 by an outside contributor and, as of 27 October, has no comments from anyone.

How to try MiroFish without burning credits

You need Node.js 18 or newer, Python 3.11 or 3.12, the uv package manager, an LLM key in OpenAI format and a Zep Cloud key. The setup file also has an optional second "boost" model slot for the high-volume calls.

git clone https://github.com/666ghj/MiroFish.git
cd MiroFish
cp .env.example .env
npm run setup:all
npm run dev

The interface opens on port 3000 and the API on 5001; docker compose up -d works instead if you prefer containers. Then:

1. Start with a short seed, a page or two, so the graph stays small.
2. Pick a cheap model for the first run; a quality model can come later.
3. Cap the rounds. The simulation script accepts a maximum-rounds setting that truncates the default 72.
4. Run the same question at least twice, ideally on two different models, and trust only what appears both times.
5. Before quoting the report, find the line in the agent logs.

If running your own stack is new territory, our playbook on self-hosted AI agents covers the trade-offs. If you are not sure which kind of tool your problem needs, Smart Match will point you at a category before you spend an evening on setup.

The case for MiroFish anyway

The best argument against this verdict: accuracy was never the point. Scenario planning has always been about surfacing possibilities a team had not considered, and a free tool that does that in an afternoon has value even if it never predicts anything.

That is fair, and it is the use this guide recommends. The objection fails only when the output is presented as a forecast, which is how the viral clips present it and how the project's own "Predict Anything" headline invites it to be read.

There is also a case on scale. OASIS is designed for casts of up to a million agents, far beyond any human panel a small team could recruit. For academic work, MiroFish is a convenient front end to a serious research tool, and the closest commercial relatives, such as Kashikoi for testing AI agents in simulated conversations, are narrower. Google Research's Vantage experiment uses simulated teammates for an even smaller job, scoring one person's collaboration.

What to watch next

Two things will change this assessment. The first is pull request 813: if a backtesting framework merges and the project publishes even one scored prediction, MiroFish moves from rehearsal tool to measurable forecaster, good or bad.

The second is the hosted edition. The mirofish.ai homepage runs a waitlist for an "online edition" with no price, and the moment it launches, the cost picture in this guide changes from your API bill to theirs.

Our prediction: by 31 March 2027, MiroFish will still not have published an accuracy figure for any real-world forecast. If it has, this page gets rewritten. You can follow the record on the MiroFish profile, and compare it with other early projects in emerging AI software.

Frequently asked questions

Is MiroFish legit?

It is real, working open-source software from a named maintainer with Shanda Group backing, so it is not a scam. What it lacks is evidence: no published simulation has been scored against a real outcome. Treat its reports as scenarios to discuss, not predictions to act on.

Can MiroFish test my website with fake customers?

No. Its agents can post, comment, like, follow and search inside two simulated social platforms, and the backend contains no browser code at all. For agents that click through a real site, look at browser agents such as Browser Use or Skyvern instead.

Can MiroFish run offline?

Not fully. The model can be any OpenAI-format endpoint, which includes servers you host yourself, but the knowledge graph and agent memory live in Zep Cloud, which the README requires. Without internet access to Zep, a run cannot build its graph.

Is the MiroFish online version available?

Not yet. The official website runs a waitlist for a hosted online edition and has not published a launch date or price. Today the only way to use it on your own material is to install the open-source version from GitHub.

Which GitHub repository is the real MiroFish?

The original lives under the 666ghj account on GitHub, maintained by the developer BaiFu. Forks with the same name exist, so check the owner before you clone, and pull the latest main branch so you get the July 2026 report-sanitising fix.

Covered in this guide

• MiroFish: Open-source app that simulates thousands of LLM personas reacting on mock social networks to rehearse public response. Self-hosted, AGPL-3.0 licensed.
• Alibaba Cloud: Alibaba Cloud, founded 2009, is China's top cloud provider (35% share) and builds the Qwen AI models, which passed 1 billion Hugging Face downloads in 2026.
• Browser Use: Open-source browser agent with a hosted cloud: describe a web task in one sentence, get JSON back. The repo has 117,000+ GitHub stars.
• Browserbase: Hosted headless browsers for AI agents, with a free plan of 3 concurrent sessions, session replays and the open source Stagehand SDK.
• CrewAI: Open-source Python framework for multi-agent AI, MIT-licensed with about 59.6K GitHub stars, plus the CrewAI AMP platform for building and governing agents.
• Eigent: Eigent is a free open-source multi-agent desktop: 4 parallel AI workers run locally on CAMEL-AI. Paid managed plans start at $19.90/month with 2,000 credits.
• Google: Google (Alphabet, NASDAQ: GOOGL) runs Search, Cloud and the Gemini models, including Gemini 4 Argon (October 2026).
• Vantage experiment: Free-to-try Google Research experiment where an AI simulated team tests and scores collaboration, creativity and critical thinking, with access by sign-up.
• Kashikoi: Kashikoi is a no-code simulation platform for evaluating AI agents, built by the ex-Moveworks lead behind 250+ shipped enterprise agents.
• Kusho: AI test automation generates production-ready API and UI test suites in minutes with 100% backend coverage and security testing.
• OpenAI: OpenAI builds ChatGPT, the GPT-6 family (Astra, Sol, Luna), Codex and the OpenAI API. It reports 1.2B weekly ChatGPT users and an $852B valuation after its March 2026 round.
• QualGent: Runs reusable test cases on iOS and Android builds and files bug reports with video, via a dashboard or an API capped at 10 requests per second.
• Skyvern: Open-source AI browser agent with 23.2k GitHub stars that logs in, fills forms, downloads files and extracts data from sites with no API.
• SuperAGI: Open-source autonomous AI agent framework and AI Super App for work, backed by $15M. Lets dev teams build, deploy, and manage multi-agent workflows with a built-in CRM and AI SDR.
• Typeform: Conversational form and survey builder founded in 2012 in Barcelona, with an AI form generator, an MCP connector and a free plan.

Sources