How it works
The whole design is that a person never has to move, name or sort a document. Here is what happens instead.
1. Arrive
Each family and business has an email-in address. Forward a bill, set the scanner to send to it, drop a file in the browser, or import the pile you already have. Forwarded mail keeps the original sender and date. Mail from a sender you have not approved waits for a yes.
2. Read
The reader (Claude, through the Anthropic API) sees the pages, the text layer, the registry of document types, and a glossary of the people and things already in your household. It records the category, the type, a title a person would use, the facts with their roles, who the document concerns, and every date it carries. Nothing on the page is invented; nothing off the page is trusted.
3. Link
"Dana L." on a prescription and "Dana Leonard" on a lease are one person. A VIN settles the car. Each document is linked to who and what it concerns, so the home screen is your household, not a folder tree.
4. Act
A term end becomes a renewal with the premium and the carrier's phone number. A lease clause becomes a notice deadline. An EOB becomes an appeal window. Each says what happens if it is missed and what to do. A newer document resolves the older one automatically.
5. Check
For each kind of person or thing, Thaw knows what should be on file and how often. The checkup lists what is missing, lapsed, expired or duplicated, worst first, with where to get it. It recomputes every time something arrives.
6. Ask
Ask in plain English. The answer is built only from excerpts of your own documents and cites the page. When nothing on file says, it says so and names the document that would.
Under the hood
The reader gets four things: the document, the extracted text layer as a cross-check, a compact summary of the type registry (the twelve categories, the kinds of paper, and the curated types with their fields), and the household glossary (the people, pets, vehicles, properties and businesses already known, with their aliases). The registry and glossary sit behind prompt-cache breakpoints, so the cost per document is the document, not the context.
It is instructed to record only what is on the page, to give a page number for every fact, to propose a new type in plain words when none fits, and to lower its confidence rather than guess. A reading below 60% confidence, a proposed type, or an unresolved person files the document with a "check this" flag. Every reading is recorded: model, prompt version, tokens, cost, duration.
Identifier-role facts (Social Security, account, policy, license, VIN and confirmation numbers) are encrypted with a blind index for exact match and redacted from page text before it is chunked and indexed. The search index is derived from redacted text only.
Early access is open to a small group of households and owner-run businesses. Tell us a little about your paper and we will send an invitation.