✦ The Desk

bring a story · manuscript → living cast
🔒 Your manuscript never leaves your machine. The file is read in this browser; extraction runs on your own local AI (LM Studio). The only thing that ever uploads — when you click publish — is the finished cast: quotes, scene summaries, character profiles. How it works ↓

1The manuscript

a chapter, an episode, a short story — plain text or markdown
Drop your story here
or click to choose a file (.txt / .md) — it stays on your machine

2Your local reader

LM Studio on this machine does the reading — nothing is sent anywhere else
If it can't connect: open LM Studio → Developer tab → start the server and switch Enable CORS on.

3The reading

chapter by chapter: scenes, dialogue, who's there — every quote checked word-for-word against your text

4Review the cast

untick anyone who's just a walk-on — you can rename before they come to life

5Bring them to life

Hypatia meets each one — builds their profile from their own words (paced politely; ~30s per character)

6Publish to The Study

the finished cast — quotes, scenes, profiles — becomes an episode you can play with

How ingestion works

  1. Local first, always. Your file is read by this page, in this browser. It is never uploaded.
  2. Your AI does the reading. The browser sends each chapter to LM Studio on your own machine (localhost), which extracts scenes, dialogue, and cast as structured data.
  3. The verbatim guard. Every extracted quote must appear word-for-word in your text (tolerant of curly quotes and dashes). Anything the model invents is rejected and counted, honestly.
  4. You review. Walk-ons out, names fixed, before anything goes anywhere.
  5. Hypatia meets them. Each kept character is introduced through the site's normal API — she builds the harmonic profile from their real words. This is the first time anything derived from your text touches the server.
  6. You publish. The episode — cast, scenes, quotes — is stored and appears in The Study. The manuscript itself remains only on your machine.

Prefer the heavier pipeline (SQLite corpus, bigger extraction passes, series continuity)? That still exists as the desk-side toolchain — see the repo's WORKFLOW doc, or just hand your file to your desk assistant.