CRATON Get a free API key →

Developers · for agents

Point your agent
at Craton.

Underwriting teams and risk managers increasingly send an agent first. Everything here is self-service over plain HTTP — a key in one call, docs that execute, results an agent can verify — and there is an MCP server with the whole build surface as tools.

Get a free API key → Read the agent walk →

Backdrop: Tokyo Bay at night from the ISS — NASA, public domain.

Three steps

Connected in minutes

  1. Add the MCP server

    Thirteen tools, from create_api_key through run_backtest to compose_model — each one a call against the public REST API, so the tools can never drift from it.

    claude mcp add --transport http craton \
  https://altier.ridgehead-hamlet.ts.net:8475/mcp

    Or in a project’s .mcp.json:

    {
  "mcpServers": {
    "craton": {
      "type": "http",
      "url": "https://altier.ridgehead-hamlet.ts.net:8475/mcp"
    }
  }
}

  2. Let it take its own key

    The create_api_key tool — or one bare call — answers with a free sandbox key, no payment details, no human in the loop. The secret prints once; the agent keeps it and sends it as a Bearer header from then on.

  3. Point it at the docs

    The documentation index is machine-readable at /llms.txt; every page serves raw markdown at /{page}.md; and the agent walk — hosts, exact request shapes, job polling — is one page: agents. The examples in those pages run against the live product on every change.

Why agents do well here

Built for a reader that checks

Results that verify

Deterministic against pinned data snapshots; a backtest re-run a year later replays byte for byte. An agent can prove its own work.

Run ids are content addresses

Derived from the document they identify — a check, not a caption. Nothing changes silently under an id that stays the same.

Assumptions as fields

Model versions, data vintages, event sets and loadings arrive beside every number, machine-readable, ready to be reasoned about.

Docs that execute

Every example is extracted verbatim and run against the live product on every change — what an agent reads is what the API does.

One error envelope

Structured refusals with a code and a message; rate-limit headers on every answer; 429 carries Retry-After. Failure is parseable.

The whole surface, as tools

Design a trigger, replay forty seasons, read a technical price, compose a model from registry primitives, start a monitor — products built end to end from a conversation.

And where it stops

Analytics, all the way down

Craton outputs technical prices, never quotes; it drafts terms, never offers; it cannot bind, issue, endorse, claim or settle. It holds no money and never touches paper — for agents exactly as for people. What an agent builds here is analysis and drafts; execution belongs to the people it works for and their own capacity.

Get a free API key →