The Contract
The Reader is useful. The way we ship it is a deliberate stance. We watched authors get burned by tools that trained on their work, by editors that turned out to be personas, by black-box assistants that nobody could audit. We do not do those things. Here is the exact mechanics, surface by surface.
Looking for the short version? What this is not is the one-page position.
I.
There is no “generate next paragraph” button. No auto-complete. No ghost text. The Reader reads your manuscript and reports back; what ends up on the page is what you typed. If you want a chatbot to draft for you, that’s what ChatGPT is for, and that is the line we will not cross. The book is yours.
II.
We do not run our own models. We do not retain your prose for any analytic purpose, we do not sell it, and there is no internal team reading manuscripts to improve features.
What happens at the provider itself depends on whose account the read ran on, and we would rather be exact than reassuring. Reads on our Anthropic account do not train anything: Anthropic does not train on API traffic. Reads on a key you connected run under the terms you agreed with that provider, and some free tiers do train on what you send them, so it is worth a look. Studio always runs on your key. Pro runs on ours, unless you have chosen “My own key” in settings and saved one, in which case it runs on yours.
III.
The Reader runs against the provider you pick: Anthropic, Gemini, or a local endpoint. A local endpoint means Ollama, LM Studio, anything OpenAI-compatible. On the Studio tier you connect your own AI account; calls bill to your account and your prose never enters our infrastructure beyond passing through.
Keys are encrypted at rest with AES-256-GCM, the same primitive your bank uses for card storage. You can clear any key from settings at any time.
IV.
Three surfaces have an inspect button: the whole-book read and character chat. The pacing and plot reads show you what was sent once the answer comes back; in character chat you tick inspect next reply before you send. The rest do not have one yet: the voice workshop, the continuity check, the grammar pass, the reverse outline, the scene synopsis, suggested chapter titles, and derived plot threads. For those, the list below is the disclosure. No black box: where you can see what we asked, you can, and where you cannot yet, we say so rather than leave you to assume.
V.
The character voice workshop only voices fictional characters you have created in your own manuscript, grounded only in scenes you wrote. We do not market “AI editors” designed to feel like real human professionals. We do not train personas on real authors’ published work. That is a line we will not cross.
VI.
In server mode, your prose is stored in a SQLite database tied to your account on Argonode infrastructure. You can export every document as Markdown at any time from your project menu. Delete your account and we wipe everything irrevocably. No tombstone, no archived copy, no soft-delete.
In local mode we never store it: the manuscript lives in your browser, on your machine, and optionally mirrors to a folder you pick. Deleting your account cannot touch that copy, because we do not hold it, and clearing it is yours to do. When you ask for a read, your browser sends the manuscript to our server, which forwards what that read needs to your provider, because that is the only way a read can happen. It passes through us. We do not keep it.
VII.
A typical Reader call is a few thousand tokens of output — roughly comparable to streaming a few seconds of video. Bigger calls (plot overview on a full novel) are correspondingly more. We disk-cache identical reads, so the second time you ask the same thing it costs zero new compute. We do not think this makes the Reader “green.” We do think a cached read is far less than the same prompt re-run in five chat tabs every morning, and we want you to make that call yourself.
VIII.
Each Reader call carries the context it needs to do its job, and no more. Some surfaces will show you the exact text after the fact; the rest are described here instead. This list covers the Reader. The optional extras that send prose elsewhere, the audiobook read, the scene visualizer and scene to video, are set out in the terms. What each read sends:
Pacing read
The project title, chapter and scene titles, word counts and dialogue ratios. No prose content is sent.
Plot read, derived plot threads
The project title, every chapter and scene title, and your manuscript in reading order. A book up to roughly 211,000 words is sent in full, whatever its number of documents. Past that, we sample passages spread evenly across the whole book by position, beginning to end, rather than stopping at the front — so the ending is always represented, however many documents the book is split into. A scene that falls between samples, or that a sample only barely clips, is left out rather than sent as a sliver too small to mean anything. On a local endpoint the whole book will not fit, so we send a much smaller sample.
Character voice workshop
The character profile fields you filled in, plus the titles and text of up to 6 scenes where the character’s name appears, totalling up to about 120,000 characters. That total is shared out by need, so a single long scene can use most of it.
Character chat and roundtable
The character profile, your message to them verbatim, and short excerpts from scenes they appear in, with those scenes’ titles.
Continuity check
The scene you are working in and the two scenes before it, with their titles.
Suggest a chapter title
Every scene title in that chapter, the opening few hundred characters of each scene, and any synopses you have written for them.
Scene synopsis, reverse outline, grammar pass
Only the one scene or passage you asked about, up to about 120,000 characters of it. The synopsis and reverse outline also send its title; the grammar pass sends the passage alone.
IX.
We treat this page as the contract. If you spot a gap, an ambiguity, or something we could explain better, email hello@argonode.studio.