Foundational models.
A registry of models for the power grid — one base model, and the price forecast products built on it. Open any card to read how it works, what it reads, what it emits, and the three ways to consume it.
Open a model card
How each model works, what it reads, what it emits, and how to consume it.
Live ERCOT price forecasts
Probabilistic price paths across ERCOT's resource nodes, load zones, and trading hubs.
Three ways to consume
In the product, over a typed inference API, or as MCP tools for your own agent.
What a model output looks like.
A Metiscast price path with its p10 to p90 band, drawn from the same inference service a developer calls.
The models.
A geotemporal time-series foundation model, and the Metiscast price forecast products built on it — live today across ERCOT’s nodes, zones, and hubs.
4 models for the physical grid. Open any one to read its full card.
Foundation model · time-series + geospatial
Our base model — a time-series foundation model pretrained on years of ISO history jointly with the grid's geography, so a forecast at one location carries what's happening around it.
02Metiscast · probabilistic price forecast · ERCOT nodal
Price forecasts at the generator- and load-level surface of the ERCOT market — the resource nodes where a specific unit actually settles, each with its published p10–p90 band.
03Metiscast · probabilistic price forecast · ERCOT zonal
Forecasts at the level where load and retail positions settle — ERCOT's eight load zones — with the same probabilistic band and issuance cadence as the nodal product.
04Metiscast · probabilistic price forecast · ERCOT hub
Forecasts at the reference points bilateral and financial positions are written against — ERCOT's seven trading hubs, including the hub average.
How the models fit together — the base model, its derivatives, and how each is reached
3
One base model
The Geotemporal Foundation Model is the backbone — a time-series foundation model pretrained on years of history from multiple ISOs, jointly with the grid's geography. The other models are built from it.
Post-trained and physical derivatives
Some models sharpen the base on a task, like the price and congestion tails. Others reason over a physical power-flow model of the network to anticipate stress and outages.
Three ways to consume each one
Every model reaches the same three surfaces: in the product, over a typed developer inference service (not MCP), and as MCP tools for your own agent.
Consume any model.
Every card reaches the same three surfaces. The model output carries its source and as-of no matter which one you reach for.
In the product
The models power Ask the Grid directly — the live map, the radar, and the grid agent read them in place.
Inference service
A direct, typed inference API for forecasts and embeddings by node and horizon. This is not MCP — it is a plain endpoint for your pipelines and services.
See the developer surfaceMCP
The same model outputs exposed as tools for your own agent. Point a compatible client at the Ask the Grid MCP server.
Build with MCPBeyond the models above
These are the models we run against the grid today. Ongoing research at the desk explores the methods behind them and what comes next.
Read the researchSee the models where the grid is live.
The agent that reads the grid lives on its own page — these models are the machinery underneath it and the research beside it.