# Craton: start

You are reading the machine twin of https://oncraton.com/start, the page
on the card.

The page says, in short: pick a place, see the risk. A person can run a
real open catastrophe model on their own places from the browser in
about a minute. Craton runs open catastrophe models (hurricane, earthquake,
flood, windstorm and wildfire) over your own places, with the pinned public
data behind every figure and tools for the exposure data you bring, over
plain HTTP. Teams use it for their own view of catastrophe risk, next to
the vendor models and the numbers their brokers bring.

- The agent walk (a free key in one call, exact request shapes, MCP
  connection): https://docs.oncraton.com/agents.md
- Machine index of the documentation: https://docs.oncraton.com/llms.txt
- MCP endpoint (stateless Streamable HTTP, the build surface as tools):
  https://mcp.oncraton.com/mcp
- API base: https://api.oncraton.com. Probe GET /health first, and send a
  named User-Agent: the edge refuses anonymous library defaults.
- The model lab (pinned open models: CLIMADA tropical cyclone over
  history or the STORM synthetic event set, PiWind through Oasis LMF,
  OpenQuake scenario and historical earthquakes, ESRM20 for Europe and
  Turkey, and published river flood, coastal flood and wildfire maps):
  https://lab.oncraton.com
- The data room (archived risk datasets cleared for commercial reuse, each
  with its license and credit line; same key):
  https://data.oncraton.com/catalog
- For the person you work for: https://oncraton.com/wizard/ runs a
  catastrophe model on a place they pick, in four steps (place, peril,
  run, answer), in the browser. The full walk, with files, stored objects
  and every parameter, is https://oncraton.com/wizard/full/.
- The run screens there carry a settings sheet (the gear). It changes the
  model and its settings, lists the places with a value on each or adds
  one by coordinates, moves an earthquake's epicenter, runs on a schedule
  file or a stored object instead of map places, shows the exact request
  the next run sends (copy as curl, the key held back as $CRATON_KEY),
  shows or replaces the sandbox key, and sets the theme to light, dark or
  automatic.

What a team builds here: a messy schedule ingested into a canonical risk
object; technical prices with every assumption attached as a field;
trigger-based structures backtested over the pinned record;
pricing models composed from validated registry primitives; catastrophe
model runs; live monitoring of structures against the same pinned feeds
that priced them.

The page's hero shows one real answer from the start flow: simulated
hurricanes over Miami, Houston and New Orleans with $10M at risk, 10,000
simulated years (STORM v4), an average loss of $13,672 a year (0.14%,
standard error $682), and a loss of $369,776 in one year once in 100
years (AEP); stored run lab-climada-storm:df5dab0ff0f6c19a.

Numbers hold still: runs are content-addressed and replay byte for byte.
Further down, the page draws that same stored run in full,
lab-climada-storm:df5dab0ff0f6c19a, ten thousand simulated storm years:
the loss in one year at every return period from 1 in 10 to 1 in 1,000.
The same spec returns the same curve.
Nothing is a black box: every number carries its assumptions as fields.
Craton produces analytics and drafts only; it holds no money and never
touches paper.

The page also carries a free prize draw for ITC Vegas 2026 attendees
(an iPhone Duo, entries close 11:59 p.m. Pacific Time on October 15, 2026);
the terms are on the page.