Point at a live world event. Get a simulation of what happens next. Get notified when it comes true.
Worldcast bridges two open-source projects — a real-time global intelligence dashboard and a multi-agent AI simulation engine — into a single tool that predicts the downstream effects of real-world events, then watches the news to tell you when those predictions actually happen.
The full loop: pick a live event → simulated prediction renders with markets / geopolitics / supply-chain bullets and watch signals → an incoming news event confirms one of those signals and a banner appears at the top. Runs against the included mocks, no Docker required — try it in 2 minutes ↓.
- Monitor — WorldMonitor aggregates 435+ live news feeds, geopolitical signals, and financial data onto a real-time 3D globe
- Select — Pick any event, click "Predict this event"
- Simulate — MiroFish spawns thousands of AI agents with memory and personality, runs a social simulation seeded by that event
- Report — Plain-English bullets across markets, geopolitics, supply chain — plus a checklist of specific verifiable signals to watch for
- Notify — The bridge continuously scans incoming WorldMonitor events against your open watch signals. When reality matches a prediction, you get a browser notification
Live event → Simulate → Report + Watch signals
↓
New events → Matcher → 🔔 "Your prediction came true"
┌──────────────────────────────────────────────────────────────────┐
│ YOUR UI :3333 │
│ Feed Tab │ Globe Tab (iframe) │ Sim Tab (iframe) │ History Tab │
│ │
│ Report Panel Watch Signals Panel │
│ ────────────── ───────────────────────── │
│ • markets □ TSMC drops >3% · 5 days 🔔 │
│ • geopolitics □ China counter-tariffs · 90d 🔔 │
│ • supply chain □ Vietnam FDI spike · Q3 🔔 │
│ [ ✓ happened ] [ ✗ didn't ] │
└──────┬───────────────────────────────────┬───────────────────────┘
│ POST /predict │ WS: progress, report,
│ │ match alerts 🔔
▼ │
┌──────────────────────────────────────────────────────────────────┐
│ BRIDGE :4000 │
│ │
│ formatter.js ── shapes WM event → MF seed │
│ jobs.js ── fires sim, polls status, streams progress │
│ summarizer.js ── raw MF report → plain English + signals │
│ watchlist.js ── stores open signals, expiry, match rules │
│ matcher.js ── NEW: polls WM feed, scores against signals │
│ notifier.js ── NEW: pushes WS alert when match found │
└──────┬────────────────────────────────────┬───────────────────────┘
│ │
▼ ▼
┌─────────────────┐ ┌─────────────────────┐
│ WorldMonitor │ │ MiroFish │
│ :5173 │◄─ matcher │ :5001 / :3000 │
│ (Docker image) │ polls │ (Docker image) │
│ │ every 60s │ │
│ GET /api/events│ │ POST /api/simulate │
└─────────────────┘ └──────────┬──────────┘
│
┌──────────▼──────────┐
│ LLM API │
│ qwen / ollama │
│ (user's own key) │
└─────────────────────┘
You only write and maintain:
bridge/— ~250 lines Node.js (6 small modules)ui/— ~180 lines HTML (single file)
WorldMonitor and MiroFish run as Docker images pulled from their registries. You never touch their code. When they update: docker compose pull.
Every 60 seconds:
matcher.js polls WorldMonitor GET /api/events (last 60s of news)
│
▼
For each open watch signal in watchlist:
score = LLM.match(incoming_event, signal_text)
if score > 0.8 → signal.status = "confirmed"
→ notifier.push(WS alert to UI)
→ UI shows 🔔 banner + marks checkbox green
if signal.expires_at < now → signal.status = "expired"
The matcher uses a lightweight LLM call — just semantic similarity between the incoming headline and the watch signal text. Cheap, fast, runs every minute.
Every prediction report ends with 3–5 specific, verifiable, time-bound signals:
WATCH FOR THESE SIGNALS status
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
□ TSMC stock drops >3% watching · expires 5 days
□ China announces counter-tariffs watching · expires 90 days
□ Vietnam FDI announcements spike watching · expires Q3
□ CHIPS Act discussion resurfaces watching · expires Q2
Last scanned: 2 min ago
When a match fires:
🔔 PREDICTION CONFIRMED
"China announces agricultural counter-tariffs"
matched incoming:
"Beijing announces retaliatory tariffs on US soybeans" · Reuters · 14 min ago
[ view event ] [ view original prediction ]
The fastest way to try Worldcast is mock mode: the bridge + UI talk to a local stub of WorldMonitor and MiroFish, so you don't need Docker, the real upstream images, a Zep Cloud account, or an LLM key beyond a local Ollama.
Requirements: Node.js 18+ and (optionally) Ollama with any small model pulled (ollama pull llama3). If Ollama isn't running, predictions just return empty summaries — everything else still works.
git clone https://github.com/wilson-cheng1110/WorldPredict
cd WorldPredict
cp .env.example .env # default values are fine for mock mode
npm install # installs bridge + test deps (~2 min)
npm run quickstart # boots mocks, bridge, UI all at onceOpen http://localhost:3333. You'll see three sample events. Click "Predict this event" on any of them — the progress bar runs through queued → building world model → summarizing → complete, then a report panel appears with Markets / Geopolitics / Supply Chain bullets and a checklist of watch signals.
To watch the notification loop fire, inject a confirming event into the mock WM:
curl -X POST http://localhost:5173/inject -H "Content-Type: application/json" -d '{
"id":"wm_demo","title":"Confirming event matches your signal",
"description":"...","source":"Demo","published_at":"2026-05-11T00:00:00Z",
"region":"TEST","category":"markets","url":"https://example.com"
}'Within 60 seconds the matcher cycle picks it up and you get a browser notification.
Ctrl-C in the quickstart terminal tears everything down.
npm test # unit tests + MF-direct tests (fast, <10s)
npm run test:integration # full loop against the running mock stackRunning against the actual WorldMonitor and MiroFish containers is not yet a one-command experience. The Docker images aren't published on a public registry, so you'd need to:
- Clone koala73/worldmonitor and 666ghj/MiroFish as siblings of this repo and build them locally
- Get a free Zep Cloud API key (MiroFish requires it for graph + simulation steps)
- Adjust
docker-compose.ymltobuild: ../worldmonitor/build: ../MiroFish/backendinstead ofimage: - Get an LLM key — Alibaba Bailian qwen-plus is free tier; OpenAI also works
See ROADMAP.md for the state of real-stack support.
# LLM — any OpenAI-compatible API
LLM_API_KEY=your_key
LLM_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
LLM_MODEL_NAME=qwen-plus
# Zep Cloud — agent long-term memory
ZEP_API_KEY=your_zep_key
# Bridge internals (defaults work, change if needed)
BRIDGE_PORT=4000
WM_URL=http://worldmonitor:5173
MF_API_URL=http://mirofish-api:5001
MF_UI_URL=http://mirofish-ui:3000
MATCH_POLL_INTERVAL_MS=60000
MATCH_CONFIDENCE_THRESHOLD=0.8
SIGNAL_DEFAULT_EXPIRY_DAYS=90worldcast/
├── package.json root npm scripts (quickstart, test)
├── docker-compose.yml real-stack composition (see ROADMAP.md)
├── .env.example all keys documented
├── scripts/
│ └── quickstart.mjs boots mocks + bridge + UI in one process
├── bridge/
│ ├── index.js Express + WebSocket server
│ ├── formatter.js WM event → MF seed document
│ ├── jobs.js fire sim, poll, stream progress
│ ├── summarizer.js raw report → bullets + watch signals
│ ├── watchlist.js store/retrieve open signals
│ ├── matcher.js poll WM feed, score vs signals
│ └── notifier.js push WS alerts to UI clients
├── ui/
│ └── index.html all 4 tabs + notification UI
├── test/
│ ├── mock-services.js stub WorldMonitor + MiroFish
│ ├── unit.test.js node:test, pure-function coverage
│ ├── mf-direct.test.js contract test against mock MF
│ └── integration.test.js POST /predict → WS → signal_confirmed
└── docs/
├── index.html GitHub Pages landing page
└── screenshots/ README assets
- Predictions are probabilistic narratives, not statistical forecasts. Scenario planning, not financial advice.
- Simulation quality scales with seed richness. Sparse events → weaker simulations.
- Each simulation takes 5–15 min and consumes significant LLM tokens. Start with <40 rounds.
- The matcher uses semantic similarity, not fact verification. A matching headline signals confirmation — human judgement still needed.
AGPL-3.0 — free to use, modify, distribute. Keep it open source.
WorldMonitor © Elie Habib — AGPL-3.0 MiroFish © 666ghj / Shanda Group — AGPL-3.0