Settings

Settings holds what is true of the application, not of one book. The game menu opens it, and the rail down the left names its pages: Appearance, Editor, AI, Layouts, Language, Shortcuts and About. Anything that belongs to the book you are writing, its mode, its targets, who is on it, is in Book settings instead.

Appearance and Editor change how the window reads. Theme (the accent tones follow it), the background star field and how much of it shows through; then the toolbar, the prose size, the text measure and the editor background. Every choice applies as you make it and is kept on this device.

Words today counts decides which words make your day. Words I wrote, the default, counts what you typed and pasted. All words added also counts imported text and text inserted from an AI answer. The choice sets the number the Session panel, the daily target, Challenges and Achievements show; the book keeps both counts either way, and both move as soon as a paste, an import or an insert lands. Both count the words that arrived: deleting text does not take words back, and an AI answer that replaces a passage counts all of its words, so neither count is how much longer the book got today.

Layouts keeps the arrangements of the panel grid. The built-in layouts and the ones you saved are listed here in menu order, and Fill gaps decides whether a neighbour grows into a region a layout leaves empty.

AI lists the lanes, the connectors and what each task runs on. The lanes are the ones this build knows, with whether each was found on this machine: a local model, Anthropic, an OpenAI-compatible endpoint, and a local harness. A connector is yours: a name, a model, an address or a command, and for the lanes that need one, a key. A key connector set to OpenAI or Other, and a custom endpoint such as LM Studio at http://127.0.0.1:1234/v1, run on the OpenAI-compatible lane. Test says whether it could be called, and the reason when it could not: unreachable, no key, no model, model not served, not installed. A local model or a custom endpoint is asked for the models it serves; a cloud key is checked without a call, so a test is never billed.

A key never goes in with the rest of the settings. It goes to this device’s secure store, the operating system’s own on the desktop and this page’s memory in a browser, and it is sent only to the endpoint of its connector. Nothing else on the page holds it, and closing the application does not send it anywhere.

The defaults table says which connector each task uses. Lookup, sweep, suggest, draft, beats, summary and the agent each take one of the connectors that tested ready, or none, in which case the task is not sent. The payload switch below shows you the bundle a task would carry before the call is made, and the cost line states the size of it.

A scene draft states how much earlier prose it carries. The scene draft dialog shows the scene’s context mode. In author-aware mode a draft always carries the last 2,000 words of prose before the scene. In strict structured mode it carries none unless you tick Include the last 2,000 words of earlier prose; the prose is then sent unchanged as raw material, and the acknowledgement you give before sending covers that choice. Changing the box drops the preparation, so the draft is prepared again before it can be sent.

Mechanical mode ignores the whole page. A book in mechanical mode makes no AI call whatever is set here, and its AI panels stand empty.

Language, Shortcuts and About. Language changes the words of the application, Shortcuts opens the key list in Help, and About names the tier this account holds and the version of the build.

Try: open Settings, AI, add a connector with a key, then look at Appearance and come back. The connector is still listed with its status, and the key field shows it masked until Reveal.