How to Use MiroFish Online — No Installation, No API Keys

You don't need to install anything to try MiroFish-style multi-agent prediction. MiroFish.us runs the full seed-to-report workflow — scenario analysis, knowledge graph, multi-agent simulation, structured prediction report — as a hosted chat you can use in the browser. Sign up, pick a credit pack (from $5.99 for 10 predictions), type your scenario, get a report. No API keys, no configuration.

This post covers both routes honestly: what self-hosting the open-source engine actually involves, and when the hosted route makes more sense.

The open-source project (and what self-hosting takes)

MiroFish went viral for good reason. The open-source repository — a "simple and universal swarm intelligence engine" — hit #1 on GitHub's global trending list and has accumulated over 17,000 stars. Under the hood it builds on OASIS, a simulation framework from CAMEL-AI that can scale social simulations to a million agents.

Running it yourself is genuinely rewarding if you're technical — and it's a real project setup, not a double-click install:

• A Python environment and the repository's dependency stack.
• LLM API keys — the agents are powered by language models, so you bring your own key and pay per token. A single rich simulation can burn through a meaningful number of calls.
• Configuration — models, agent counts, simulation parameters.
• Somewhere to run it — your machine for small runs, or a VPS for anything serious. Hosting providers now offer one-click MiroFish VPS templates, which solves deployment but still leaves you managing keys, updates, and costs.

If that list sounds like a fun weekend, the repo is excellent and you should absolutely star it. If it sounds like three evenings you don't have, keep reading.

The browser route: prediction in two minutes

MiroFish.us packages the same workflow — the one described in our complete guide — behind a chat interface:

1. 1.Create an account and pick a credit pack — from $5.99 for 10 predictions. No API keys to manage, and credits never expire.
2. 2.Describe your scenario in plain language. A product launch, a pricing change, a policy shift — here's how to write a prompt that gets a sharp report.
3. 3.Watch the pipeline run — seed analysis, graph building, the simulation rounds, then report generation, with live progress for each stage.
4. 4.Read and interrogate the report — most-likely outcome with probability, alternative scenarios, risks, and recommended actions, plus a follow-up chat grounded in your report.

No environment variables were harmed in the making of that prediction.

Self-host or hosted? An honest comparison

• Choose self-hosting if you want to modify the engine, run massive agent counts, keep everything on your own infrastructure, or learn how the internals work. It's open source — that freedom is the point.
• Choose the hosted version if you want answers, not infrastructure: no setup, no per-token API billing, reports in minutes, chat follow-ups, and history saved to your account.
• Cost reality check: self-hosting is "free" the way a free puppy is free — the VPS and LLM API usage are yours to feed. Hosted credits start at $5.99 for 10 predictions, which is typically less than the raw API cost of running equivalent simulations yourself, before counting your time.
• Privacy: hosted predictions are private to your account. If your scenario is classified-level sensitive, self-host on your own metal — that's the right call, and we'll say so.

What to run first

New users get the best first impression from a scenario with real stakeholders and a real tension — not "will it rain tomorrow." Steal one of these ten use cases with copy-paste prompts, or start from the decision you're actually facing this week: the pricing change you're debating, the launch you're nervous about, the announcement you keep postponing.

The fastest way to evaluate any prediction engine — hosted or self-hosted — is to ask it about a decision where you already know the stakes.

Run your first prediction — and if you do go the self-host route instead, take our prompt-writing rules with you. They work on any engine.