MiroFish Setup Guide 2026: Run Swarm Prediction on Remote M4 Mac

What is MiroFish?

MiroFish (homepage: mirofish.ai, 60K+ GitHub stars as of May 2026) is not a single LLM chatbot. It is a five-stage prediction pipeline:

1. Graph building — seed extraction, individual/collective memory, GraphRAG
2. Environment setup — entity graphs, persona generation, agent injection
3. Simulation — dual-platform parallel runs, temporal memory updates (powered by OASIS)
4. Report generation — ReportAgent with tool access to post-sim state
5. Deep interaction — chat with simulated agents or the report agent

Use cases published by the community include public-opinion forecasting, policy what-if analysis, and prediction-market rehearsal (always validate with human review — emergent sims are not ground truth).

How it differs from TradingAgents vs FinGPT: TradingAgents models a trading-firm committee (analyst → risk → BUY/HOLD/SELL). MiroFish models social swarms (thousands of personas interacting). Many quant teams run TradingAgents for execution logic and MiroFish for scenario rehearsal before capital deployment.

Hardware and RAM: What a Remote M4 Needs

MiroFish bills itself as scaling to very large agent counts; your bottleneck is LLM API throughput and RAM for orchestration, not Metal GPU.

Rule of thumb: reserve ~2–4 GB for Docker, ~4–8 GB for Python orchestration, and headroom for in-memory graph structures. If RSS exceeds 14 GB on a 16 GB node, upgrade per M4 16GB vs 24GB matrix.

For rental window planning (3-day spike vs monthly lab), see burst vs monthly guide.

MiroFish Setup on KuzCloud: Step by Step

Step 1 — Provision node and SSH in

1. Order an M4 node (Japan or Hong Kong for Asia API latency; US East for OpenAI-native keys).
2. SSH with issued key: ssh -i key.pem user@node-ip
3. Install Docker Desktop CLI or Colima if not preinstalled: brew install colima docker && colima start

Credentials typically arrive within ~5 minutes of payment — same flow as OpenClaw remote setup.

Step 2 — Clone and configure environment

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

Edit .env with your LLM provider keys (project supports multiple backends via CAMEL/OASIS stack). Never commit .env to git.

Step 3 — Docker Compose (recommended)

docker compose up -d

Default ports (verify in repo docker-compose.yml for your release tag):

Access UI via SSH tunnel if the node has no public ingress:

ssh -L 3000:localhost:3000 -L 5001:localhost:5001 -i key.pem user@node-ip

Then open http://localhost:3000 on your laptop.

Step 4 — Run a prediction workflow

1. Upload seed materials (PDF, news dump, or structured market brief).
2. Describe the prediction goal in natural language (e.g., "How will retail sentiment shift if rate cuts delay 90 days?").
3. Wait for graph build → environment setup → simulation → report.
4. Use Deep Interaction to query individual agents or ReportAgent.

First full run on 16 GB often takes 30–90 minutes depending on agent count and API rate limits.

Step 5 — Persist artifacts and tear down

Store reports under ~/mirofish-runs/ on the node NVMe. When finished, docker compose down and release the node if on a daily burst plan to avoid idle billing.

Network, API, and Compliance

• LLM costs dominate: 1,000 agents × multiple turns can exceed $50–200 per experiment with frontier models — model smaller personas for iteration.
• AGPL-3.0: Network use may trigger source-sharing obligations if you offer MiroFish as a service — read the license before production SaaS.
• Data residency: Seed documents live on the rented Mac disk; pick region nodes matching your compliance needs (region matrix).
• Not live trading: Pair with TradingAgents vs FinGPT for separate execution-layer logic; MiroFish does not replace exchange connectivity.

External references: MiroFish GitHub, CAMEL-AI OASIS.

Troubleshooting

For general Mac rental economics, see Mac Mini rent vs buy.

FAQ

Can MiroFish run without Docker on macOS?

Yes — Python 3.10+ and Node/Vue builds are documented upstream, but Docker is the most reproducible path on a fresh rented Mac.

Does MiroFish need Apple Silicon GPU?

No. Inference is API-based or CPU-bound orchestration. GPU helps only if you wire local Ollama models through compatible backends.

How is MiroFish different from OpenClaw?

OpenClaw is a coding-agent daemon for repos and CI. MiroFish is a social simulation engine for prediction reports — complementary tools.

What is the minimum rental period to evaluate MiroFish?

A 3–7 day burst suffices for Docker setup, one seed run, and report review if you parallelize API calls during business hours.

Can I run MiroFish and TradingAgents on one M4?

Yes on 24 GB: TradingAgents orchestration (~2–5 GB) plus MiroFish Docker (~4 GB) leaves room for moderate agent counts. Do not run both heavy sims simultaneously.