jangles-byte/Pythia

Pythia is a local forecasting engine that fuses dozens of live, free data feeds into a single world-state and then asks a local LLM — running entirely on your own hardware via Ollama — to predict what happens next across four time horizons: 24 hours, a week, a month, and a year. Every prediction carries a probability, a geographic anchor, and the reasoning behind it. Click a forecast and the 3D globe flies to the location.

The architecture is a three-layer stack. Osiris provides the live globe and feed layer — breaking news, GDELT geopolitics, armed conflict, NWS storm polygons, EONET disasters, wildfires, earthquakes, cyber threats, and more. The PYTHIA engine pulls all of that concurrently, fuses it into a prose world brief, and hands it to a local LLM for drafting. Then a council of four specialist personas — Strategist, Economist, Naturalist, Skeptic — each re-scores every forecast through its own lens. The system surfaces not just the consensus number but where the swarm splits, which is the more honest output.

What makes this more than a dashboard is the accountability layer. Every forecast is written to a ledger the moment it is made. When a horizon expires, an LLM judge grades the prediction against the archived world state and updates a Brier score — broken down per horizon, per swarm persona, and per model. The consensus vote is Brier-weighted, so a persona that has been consistently right earns a louder voice over time, clamped so no single agent dominates. That feedback loop is the thing most forecasting tools skip entirely.

The feed count is genuinely wide: 48 sources by the README's own count, including SEC EDGAR insider trades (Form 4 buys and sells across all public companies), Polymarket and Kalshi event-contract odds as forecasting anchors, CDC wastewater surveillance that leads clinical case counts by roughly one to two weeks, NOAA ENSO indices, OFAC sanctions actions, IODA internet outage data, and live power-grid carbon intensity for both the GB National Grid and California CAISO. None of these require API keys or accounts.

For agent builders, the pitch is concrete. A single GET /agent/view call returns a prose world summary, every live event grouped by domain with coordinates, and the current predictions — one payload that gives an agent situational awareness it otherwise lacks entirely. The full API runs on localhost port 8088 with CORS open, no rate limits, and an MCP server so Claude Code or any MCP client can treat PYTHIA's tools — world_brief, get_predictions, what_if, get_scorecard — as native capabilities.

The honest caveat is that forecast quality depends entirely on whatever Ollama model you pull locally. A small model will produce shallow reasoning; the swarm deliberation and Brier weighting help calibrate over time, but they cannot compensate for a weak base model. The counterfactual (/whatif) endpoint is explicitly ephemeral and never touches the track record, which is the right call — mixing hypotheticals into the scorecard would corrupt the only honest signal the system has.