Multi-Agent AI Prediction Engine — Fully Self-Hosted with Neo4j
A fork of MiroFish that replaces Zep Cloud with Neo4j — no vendor lock-in, no cloud dependency, fully offline.
The original MiroFish depends on Zep Cloud — a closed-source, cloud-hosted graph database. This fork replaces it with a local Neo4j instance:
| Zep Cloud (Original) | Neo4j (This Fork) | |
|---|---|---|
| Hosting | Cloud-only | Local Docker container |
| Cost | Free tier limits | Completely free |
| Privacy | Data sent to cloud | Data stays on your machine |
| Offline | Requires internet | Works fully offline |
| Open Source | Proprietary | Neo4j Community Edition |
| Query Language | Custom API | Cypher (industry standard) |
| Browser UI | Limited | Full Neo4j Browser at :7474 |
MiroFish prediction engine — upload documents and build knowledge graphs
Ontology generation + graph construction from uploaded documents
Interactive graph visualization with entities and relationships
AI-generated agent personas from the knowledge graph
Custom agent configuration before running the simulation
Parallel multi-agent simulation with real-time interaction view
Comprehensive report with graph view and interactive agent Q&A
- Docker & Docker Compose
- An OpenAI API key (for LLM reasoning)
cp .env.example .env
# Edit .env — add your OpenAI API keydocker compose up -d| Service | URL | Credentials |
|---|---|---|
| MiroFish UI | http://localhost:3000 | — |
| MiroFish API | http://localhost:5001 | — |
| Neo4j Browser | http://localhost:7474 | neo4j / mirofish2026 |
| ChromaDB | http://localhost:8000 | — |
| Whisper ASR | http://localhost:9000 | — |
All Neo4j patches are automatically applied on every docker compose up via a custom entrypoint script. No manual steps needed.
docker compose up -d
│
▼
┌──────────────────┐
│ patches/ │ Auto-applied at startup:
│ entrypoint.sh │ • graph_builder.py → Neo4j
│ │ • zep_entity_reader.py → Neo4j
│ │ • zep_tools.py → Neo4j monkey-patch
│ │ • config.py → skip Zep validation
└──────┬───────────┘
▼
┌──────────────┐ ┌────────────┐ ┌───────────┐ ┌─────────┐
│ MiroFish │ │ Neo4j │ │ ChromaDB │ │ Whisper │
│ :3000/:5001 │ │ :7474/:7687│ │ :8000 │ │ :9000 │
└──────────────┘ └────────────┘ └───────────┘ └─────────┘
.
├── .env.example # Environment template (copy to .env)
├── .gitignore # Keeps secrets and data out of git
├── docker-compose.yml # All services + auto-patching entrypoint
├── README.md
│
├── images/ # Screenshots for documentation
│
└── patches/ # Neo4j patches (auto-applied on startup)
├── entrypoint.sh # Startup script that applies all patches
├── neo4j_graph_builder.py # Replaces Zep graph builder with Neo4j
├── neo4j_entity_reader.py # Replaces Zep entity reader with Neo4j
├── neo4j_graph_api.py # Flask blueprint for /api/neo4j/ endpoints
└── patch_zep_tools.py # Monkey-patches ZepToolsService → Neo4j
| Variable | Description | Required |
|---|---|---|
LLM_API_KEY |
OpenAI API key (or compatible) | Yes |
LLM_BASE_URL |
LLM API base URL | Yes |
LLM_MODEL_NAME |
Model name (e.g. gpt-4o) |
Yes |
ZEP_API_KEY |
Placeholder (needed by original image to start) | Yes* |
NEO4J_URI |
Neo4j Bolt URI | No (default in compose) |
NEO4J_PASSWORD |
Neo4j password | No (default: mirofish2026) |
| Feature | Status |
|---|---|
| MiroFish UI at localhost:3000 | ✅ Working |
| LLM-powered ontology generation | ✅ Working |
| Knowledge graph building → Neo4j | ✅ Working |
| Neo4j Browser at localhost:7474 | ✅ Working |
| ChromaDB vector embeddings | ✅ Working |
| Whisper audio transcription | ✅ Working |
| Entity reading from Neo4j | ✅ Working |
| Agent persona generation | ✅ Working |
| Multi-agent simulation | ✅ Working |
| Final report + interactive Q&A | ✅ Working |
Auto-patching on docker compose up |
✅ Working |
MiroFish Neo4j Fork is not just a research tool — it is a production-grade infrastructure for building intelligent prediction systems. The combination of multi-agent simulation, knowledge graphs, and parallel consensus makes it uniquely suited for high-stakes forecasting environments.
This engine can serve as the decision-making backbone for prediction market platforms like Polymarket, Kalshi, and decentralized forecasting protocols:
- Autonomous Market Agents — Deploy specialized AI agents that independently research, reason, and form probabilistic forecasts on real-world events (elections, macro indicators, geopolitical outcomes).
- Bot Trading Pipelines — Pipe agent consensus outputs directly into trading strategies. Each agent acts as an independent analyst; the aggregated signal drives position sizing and market entry/exit.
- Event-Driven Forecasting — Upload breaking news, earnings reports, or policy documents. The system builds a knowledge graph in real time and spawns agents to evaluate downstream effects — enabling rapid, evidence-based market positioning.
Unlike single-model inference, MiroFish runs parallel multi-agent simulations where each agent operates with a distinct persona, knowledge base, and reasoning strategy:
- Diverse Cognitive Profiles — Agents are generated from the knowledge graph with unique expertise, biases, and analytical lenses. This mirrors ensemble methods in ML — reducing variance and improving prediction robustness.
- Structured Deliberation — Agents interact, challenge each other's reasoning, and converge on consensus through structured debate rounds. The final output is not a single opinion but a calibrated collective forecast.
- Consensus Confidence Scoring — The degree of agent agreement provides a built-in confidence metric. High consensus = high-conviction trade signal. Divergence = uncertainty flag.
| Capability | Status | Description |
|---|---|---|
| Knowledge graph from unstructured data | ✅ Live | Ingest documents, news, transcripts → structured graph |
| Multi-agent persona generation | ✅ Live | Graph-aware agents with domain expertise |
| Parallel agent simulation & debate | ✅ Live | Independent reasoning + structured consensus |
| Final consensus report with confidence | ✅ Live | Aggregated prediction with supporting evidence |
| Prediction market API integration | 🔜 Roadmap | Direct integration with Polymarket / Kalshi APIs |
| Real-time event stream ingestion | 🔜 Roadmap | Live news feeds → continuous graph updates → agent re-evaluation |
| Backtesting framework | 🔜 Roadmap | Historical event replay to validate agent accuracy |
| Portfolio-level signal aggregation | 🔜 Roadmap | Multi-event, multi-market position management |
| On-chain agent wallets | 🔜 Roadmap | Autonomous agents with signing capability for DeFi prediction markets |
- Self-hosted & private — Your trading signals, data sources, and agent strategies never leave your infrastructure. Zero data leakage to third-party APIs (beyond LLM calls).
- Knowledge graph advantage — Neo4j provides structured relational reasoning that flat-context LLMs cannot. Agents reason over entity relationships, causal chains, and temporal dependencies.
- Scalable consensus — Spin up 5 agents or 50. More agents = richer deliberation = more robust predictions. The architecture scales horizontally.
- Auditable reasoning — Every agent's reasoning chain is logged and queryable. Full transparency into why the system made a particular prediction — critical for regulatory compliance and strategy iteration.
- Model-agnostic — Swap GPT-4o for Claude, Llama, Mistral, or any OpenAI-compatible endpoint. Mix models across agents for cognitive diversity.
TL;DR — MiroFish Neo4j Fork transforms multi-agent AI simulation into an actionable prediction engine. Feed it information, let agents deliberate in parallel, extract consensus forecasts, and use those signals to trade prediction markets with conviction.
Fork of MiroFish by the MiroFish team. Simulation engine powered by OASIS from CAMEL-AI.
Infrastructure: Neo4j · ChromaDB · OpenAI Whisper · OpenAI GPT
Based on MiroFish — licensed under AGPL-3.0.
Built by @0xrphl