P&C insurance · open source · local-first

Turn a rate filing into a working rating engine. Every number, cited.

P&C rating manuals are hundreds of pages of PDF. OpenRater makes them executable: an AI transcribes the filing into an Excel workbook you review, a deterministic engine rates, and every factor traces to its page.¹

↓ Download View on GitHub

free · Apple silicon · Intel Mac · Windows

OpenRater plan report: a Businessowners plan for Nebraska showing a $826 reference premium, the rating chain, and ranked price drivers — Fig. 1 · The plan report — the algorithm narrated, price drivers ranked.

40/40 filing examples reprice exactly every cell cited to a page 0 AI calls at rating time 127.0.0.1 only · no telemetry

§ 1

How it works

filing.pdf → workbook.xlsx → plan + build report → premium

01filing.pdf

You bring the manual you're entitled to use. OpenRater never fetches filings.

you

02workbook.xlsx

The AI transcribes it into a spec'd workbook — every cell cites its page, and anything unsupported lands in a gaps sheet, never approximated. You audit it in Excel.

ai transcribes · you review

03plan + report

A deterministic compiler validates and builds the plan — zero AI from here on. The filing's own worked examples run as acceptance tests.

platform

04premium

Walk any risk step by step, quote in chat, serve a local API, or re-rate a book from CSV. Same inputs, same answer, every time.

platform

§ 2

Why trust it

Every number has a citation

The build report records where each factor came from, to the page. Revisions show a cell-grain diff before anything is applied.

The filing tests itself

Worked examples from the source document become acceptance tests in the test_cases sheet — correctness is pinned to the filing, not to opinion.

Refuse, never improvise

An input the plan can't rate is a visible, structured error — never a silent neutral factor, never a plausible-but-wrong premium.

A human seat, built in

The workbook is the review gate: an ordinary Excel file an actuary can open, audit, and correct before anything rates.

The build report says so — or it says no.

Every build replays the filing's worked examples through the production engine: 40 of 40 checks match to the penny, warnings are itemized, and every transcriber assumption is flagged for review — before the first real risk is rated.

OpenRater build report: 40 of 40 filing-example checks matching, plan checks itemized, and transcriber assumptions flagged for review — Fig. 2 · The build report — verified against the filing's own examples.

§ 3

Install in two minutes

Get Claude Desktop

Download Claude Desktop (Mac or Windows) and sign in. It's the only prerequisite.

Download one file

Pick yours:

Drag it in

In Claude Desktop, open Settings → Extensions and drag the file in. Signed and notarized — no warnings.

Ask Claude

Start a chat and ask:"What is OpenRater? Show me." The engine starts locally, the app opens in your browser, and a three-minute tour begins.

Not sure which Mac? Apple menu → About This Mac — "Chip" means Apple silicon, "Processor" means Intel. macOS builds are Developer ID-signed and notarized; Windows builds are signed via Azure Trusted Signing.

§ 4

Three ways in

actuary

Claude Desktop + the extension. The full loop in chat, review UI one click away. No terminal required.

engineer

Claude Code, or any MCP client. The same loop, scriptable, over 18 MCP tools — see AGENTS.md.

integrator

Self-hosted. REST API, deterministic scoring engine, integration-events ledger — a Docker Compose kit ships in deploy/.