Turning on AI

AI in this app is off until you turn it on, and even then it's a tool you point at your prose, not a service that runs on its own. You choose the provider — a cloud model, or a local one through Ollama with the network switched off — and you see the cost of every call. This guide gets it set up.

Turning AI on is two deliberate steps, plus an optional third:

1. Set a policy — permission for a project to reach AI at all.
2. Add a provider — the API key (or local host) the models run through.
3. Hire assistants — optional named "who runs this" configurations.

Once those are done, you drive the AI through prompts, chats, and roleplay — each its own guide (Writing prompts, Context picker, Roleplay).

Step 1 — Set the policy

Policy is the permission layer. It fails closed: nothing reaches a provider unless a policy allows it, and widening access is always an explicit save.

For a project: open the project (its own window), then click AI Policy. The AI access dialog offers:

• Inherit — take the policy from a parent project, or Off at the top of the chain.
• Off — no AI at all.
• Local only — only local models (Ollama); nothing leaves your machine.
• Cloud allowed — cloud providers permitted (Anthropic, OpenAI, OpenRouter).

Pick one and click Save — a selection alone never takes effect until you do.

App-wide default: in ≡ menu → Settings… → AI, the Default AI access section sets the policy a project falls back to when it states none of its own. It has its own Apply button, so raising the floor is a separate, deliberate act.

Step 2 — Add a provider

Open ≡ menu → Settings… → AI.

Cloud providers — under the provider list, click + Add provider, choose Anthropic, OpenAI, or OpenRouter, paste its API key, and Save. Configured providers show as chips; you can rotate a key or remove one later. Keys are stored in your system keychain (Windows Credential Manager, macOS Keychain, or the Linux Secret Service), never in your project files or in the app's config.yaml — so a project folder is safe to share, and so is the config file. On a machine with no keychain (e.g. a headless Linux server) keys fall back to config.yaml, and Settings says so.

The app never asks you to type a key anywhere but here, and keys are masked on read. If you'd rather not use the cloud at all, skip straight to Ollama.

Local models (Ollama) — set the Ollama host field (default http://127.0.0.1:11434). With Ollama running and a project policy of Local only, the whole app works offline. How the app keeps a local model fast — sizing its context window so a chat doesn't spill onto the CPU, and the one Ollama setting worth tuning — is its own guide: Ollama context.

Check it works — the Connection section has a Test connection button. If it's disabled, it tells you why (e.g. "This project's AI access is off, so there is nothing to reach.").

Step 3 — Hire an assistant

Why assistant, and why hire? Because that's the role the app gives AI — a helper you keep beside your desk, not a faceless engine. You subscribe to a provider (that was Step 2), then hire an assistant: give it a name and a model and point your work at it. It's the app's whole stance in a word — a research assistant and a sparring partner next to your manuscript, never a ghostwriter that owns it. The vocabulary is deliberate: you staff your book, rather than configure a model.

So an assistant is, concretely, a reusable "who runs this" — a name paired with a provider and a model. Instead of choosing a model for every call, you set up an assistant once and point work at it.

• Create one from ≡ menu → Assistants (a project must be open) with the pane's + button. The create-project wizard also offers this up front, under Assistants for this book → + Hire an assistant….
• Choose its model in the assistant's Details rail: pick a Subscription (provider), then a Capability tier — ⚡ Fast, ⚖ Balanced, ✨ Premium, 🧠 Reasoning, or 💻 Local (Ollama) — and, under Advanced, an exact Model if you want to pin one.
• Curate the roster — the Assistants pane lists your assistants (each showing its provider · model); List / Un-list and drag to order them. The topmost is the default.

Thinking in tiers rather than exact model names means you can say "use a fast model here, a premium one there" and let each provider's current best fill the slot.

What it costs

Every AI call is metered. Costs show in the editor as you work — per call and as a running project total — so nothing is hidden. (Amounts display in euros.) The per-character and per-scene breakdowns are covered in the Roleplay guide.

You're set

With a policy, a provider, and at least one assistant, AI is ready. From here:

• Writing prompts — the templates that turn your project into structured requests.
• Context picker — hand a prompt exactly the scenes and lore it should see.
• Roleplay — let characters take turns in a scene.

AI stays off everywhere you don't switch it on, sees only the context you give it, and — with Ollama — never has to touch the network at all.