Novelcrafter approaches AI fiction writing differently from conventional all-in-one generative tools. Rather than locking writers into a single proprietary model or bundling steep token markups into a monthly subscription, Novelcrafter operates on a Bring Your Own Key (BYOK) architecture. The platform provides the structural scaffolding—the distraction-free manuscript editor, scene beat workflows, and the Codex story bible—while allowing novelists to connect directly to commercial providers, open-source aggregators, or local offline inference engines.
To harness this setup effectively, novelists need to complete three foundational tasks: connect their API credentials, assign models to prompt presets, and calibrate generative sampling parameters (such as temperature, top P, and repetition penalties) to match their desired prose voice.
1. Connecting AI Providers and API Keys
Novelcrafter manages AI connections at the account level, accessible through Profile Image → Open Settings → AI Credentials (or AI Connections). Crucially, AI credentials are stored client-side in the local browser storage rather than synced across cloud databases. If you clear browser caches, write in private/incognito windows, or switch between workstations, you will need to reconnect your keys or export and re-import your connection configuration.
Option A: OpenRouter (Recommended for Most Authors)
For novelists who want access to a wide variety of models without juggling separate billing accounts, OpenRouter is the most flexible integration:
- In the AI Credentials menu, locate OpenRouter and click Connect.
- Select Continue to OpenRouter, authenticate your account, and approve the connection.
- You can establish an optional per-key spending limit during authorization. If you set a budget limit and subsequently hit it, Novelcrafter will flag an invalid key or insufficient credit error until you update the threshold in your OpenRouter account dashboard.
- OpenRouter unlocks flagship models like Anthropic's Claude series, OpenAI's GPT models, Mistral, and community fiction fine-tunes under a single prepaid credit balance.
Option B: Direct OpenAI Connections
If you have a dedicated developer account with OpenAI:
- Navigate to your OpenAI developer console to generate an API secret key. Note that an OpenAI API account is distinct from a consumer ChatGPT Plus subscription; ChatGPT Plus does not grant API token access.
- Under Novelcrafter's AI Credentials, select OpenAI, paste your API key, and save.
Option C: OpenAI-Compatible Endpoints (Custom Clouds & Local Models)
Novelcrafter also supports third-party inference providers and local engines via the standardized OpenAI API format:
- Scroll to the bottom of the AI Credentials menu and select OpenAI Compatible Connection → Add.
- Enter a descriptive provider name (such as Fireworks, Grok / xAI, Groq, Mancer, NanoGPT, or NovitaAI) and the provider's Base URL (which typically ends in
/v1, such ashttps://api.groq.com/openai/v1orhttps://api.x.ai/v1). - Paste the corresponding API key. Novelcrafter will automatically query the endpoint to detect available models. If automatic detection fails, enable Use custom model list and enter the exact model identifiers separated by commas or line breaks.
- Local Offline Inference: Authors running local models on desktop hardware via tools like LM Studio or Ollama can enter their local localhost endpoint (for example,
http://localhost:1234/v1). This enables offline, zero-data-leak drafting with no per-token billing. - Direct Anthropic Note: Due to browser cross-origin resource sharing (CORS) constraints, the Anthropic Claude developer console cannot be pinged directly from a browser-based application without an intermediary. Authors wishing to use Claude should route requests through OpenRouter's BYOK tier or run a local proxy such as LiteLLM.
2. Managing Models and Custom Prompt Presets
Once credentials are connected, models are mapped to writing actions inside the Prompt Library rather than chosen globally. Novelcrafter separates AI tasks into dedicated prompts: Scene Beat Completion, Continue Writing, Scene Summarization, and Chat Assistants.
To customize the models available for a specific task:
- Open the Prompt Library from the main dashboard or the project left-hand sidebar.
- Select the prompt you wish to configure (such as the default General Purpose Scene Beat Completion prompt).
- If using a read-only system prompt, click Clone to create an editable custom duplicate, as system defaults cannot be altered directly.
- In the prompt editor, navigate to the General tab and select Models.
- Here you can add individual models, remove unused options, or attach curated Model Collections. Grouping models by function—such as "Brainstorming Fast" (e.g., Groq Llama 3) versus "Deep Prose" (e.g., Claude 3.5 Sonnet)—ensures quick switching during active drafting.
3. Calibrating Prose Generation Settings
In fiction drafting, raw model capability matters far less than sampling calibration. Clicking the expand arrow (>) next to any model in your prompt settings reveals the granular parameters that dictate style, variance, and rhythm.
Temperature: Creativity vs. Fidelity
- The Mechanism: Controls the probability distribution of generated tokens.
- When to Lower (0.3 – 0.6): Use lower temperature for summarizing completed scenes, maintaining rigid character speech patterns, or executing plot beats that must follow specific clues without diverging.
- When to Raise (0.7 – 1.0+): Use higher temperature when prose feels repetitive, robotic, or overly predictable, or when drafting sensory-rich scene descriptions, surreal sequences, or high-concept brainstorms.
Top P: Controlling Diction and Outliers
- The Mechanism: Restricts the model's candidate tokens to the top cumulative probability percentage.
- Practical Fiction Use: If you notice a model hallucinating nonsensical words, malformed compound adjectives, or bizarre punctuation loops (a common issue when pushing high temperatures on open-source models), lower Top P to 0.85 – 0.90 to restrict choices to safer lexical pools. Raising Top P expands vocabulary diversity.
Max Tokens: Output Length and Pacing
- The Mechanism: Imposes an absolute token cutoff on model generation.
- Practical Fiction Use: Because language models stop abruptly mid-sentence when they reach their max token threshold, avoid using low token caps to dictate scene length. Set Max Tokens high enough to accommodate a complete beat (e.g., 800–1,500 tokens), and instead guide desired prose length using structured prompt instructions or Novelcrafter's prompt input variables.
Penalties: Combating Repetition and Loops
- Frequency Penalty: Penalizes words proportionally to how many times they have already appeared in the current output. Increase this slightly (e.g., +0.1 to +0.3) if the model fixates on pet words (such as "suddenly," "echoed," or "gaze"). Avoid setting it too high, or the model may begin dodging character names.
- Presence Penalty: Penalizes the re-introduction of concepts and topics once mentioned. Use this when drafting dialogue or scene transitions where the model keeps circling back to the same argument.
- Repetition Penalty: Standard across local and open-source models (typically ranging from 1.05 to 1.15) to halt repetitive phrase looping.
- Reasoning / Thinking Effort: For hybrid reasoning models (such as Claude thinking modes or OpenAI reasoning variants), configure reasoning parameters to allow the model to plan scene beats and verify continuity against Codex lore before writing the actual prose.
4. Drafting with Custom Settings in the Write Interface
Once your credentials and prompt presets are established, using them in your manuscript is seamless:
- Open any scene in the Write view.
- In the body of the editor, type
/to open the command slash menu and select Scene Beat. - Enter your beat instructions (what happens in the beat, emotional beats, or character actions). Novelcrafter automatically parses referenced Codex entities (characters, locations, items).
- Click the dropdown menu next to the Play (▶) button to select your configured model preset, or choose Tweak and Generate to adjust contextual inputs on the fly.
- After the model responds, evaluate the prose:
- Apply: Inserts the generated text directly into the manuscript body.
- Retry: Discards the generation and queries the model again with the same parameters.
- Discard: Cancels the output without affecting the manuscript.
- Section: Wraps the output into a collapsible block, allowing you to generate multiple alternative drafts from different models and kitbash the strongest sentences into your final prose.
By combining browser-stored BYOK connections with tailored prompt presets, novelists gain total transparency over AI expenses, eliminate vendor lock-in, and gain precise editorial control over every chapter.