Operator can read documents properly, on your own machine, with no model tokens spent on the reading itself.

What “properly” means

Not “paste the text into a prompt”. The document runtime extracts structure: headings, tables, figures, and where on the page each thing came from. An answer can point at page 412 and the box on it, and you can check.

There is an OCR ladder for scans, so a photographed page still becomes text.

This runs as a plain program in the box, not as an agent. No prompt, no model call, no tokens. The consequence that matters: your documents never leave your Mac and are never sent to a model provider unless a later step you asked for sends an excerpt.

Getting documents in

• Attach them to a job on the launch sheet.
• Drop them in a watched folder and let an automation pick them up.
• Send one from your phone straight into a job’s folder on the Mac, with no cloud round trip.

Knowledge bases

A set of documents can become a knowledge base an agent can search and cite. Jobs that use one give answers with references back to the source document and page.

Say “cited answers from my own documents” rather than expecting it to understand your business. It is very good at finding and quoting the right passage, which is most of what you actually want.

The first thing to try

Take the longest, most annoying PDF you have. A contract, a manual, a spec. Attach it to a job that reads documents and ask a specific question with a checkable answer.

Then check the citation. That one check is what tells you whether to trust it on the next hundred pages, and it takes a minute.

Where the output goes

Structured results land in the session as documents you can open, share, or hand to another job. The originals are untouched.