NovelGenerator — local AI novel generator for Gemini and Ollama

NovelGenerator is a source-available AI novel writer that turns a one-paragraph premise into a full manuscript, chapter by chapter. Run it with Gemini or use Ollama as a local AI writer with your own model.

It is built for long-form story generation rather than isolated prompts: causal planning, persistent story memory, explicit scene handoffs and continuity checks keep characters, facts and open plot threads available as the novel grows.

AI novel generator · Local LLM and Ollama · Long-form fiction · Persistent story memory · Continuity checking

Run the AI novel generator locally

git clone https://github.com/KazKozDev/NovelGenerator.git

cd NovelGenerator
npm install
npm run dev

  VITE v6.3.6  ready in 118 ms

  ➜  Local:   http://localhost:3000/
  ➜  Network: http://192.168.1.177:3000/

Open that address, pick a provider, set a model, write a premise and a chapter count, then start. Gemini needs GEMINI_API_KEY in the environment; Ollama needs to be running. The engine designs the book, reviews its own construction, then writes chapter by chapter and scene by scene without asking you to approve anything mid-run. Export the finished book as EPUB, PDF, TXT or Markdown, and the audit as JSON.

Use Ollama as a local AI novel writer

Select Ollama in the app to write with a model served from your machine. This is the local/private path: manuscript generation does not require sending the premise or chapters to Gemini. You can use one model for prose and a separate Ollama model for planning, continuity review and the final audit.

Write a full novel chapter by chapter

The premise is the whole input. From it the engine builds one compact construction — cast, world rules, a causal map, an ending and a chapter map — and refuses to write against a construction its own review found incoherent. A premise name nobody in the cast answers to fails that review in code, not by opinion.

A measured run: 7 chapters, 15,390 words, 122 model calls, about 40 minutes on Ollama cloud models.

Keep continuity without rereading the book

Every scene is folded into story memory before the next one is written. One call extracts what changed — events, locations, who knows what, which promises opened and closed — and each record must cite the paragraph that proves it; a citation pointing nowhere is rejected and asked again.

The accepted scene produces an explicit handoff: what the reader already knows, what actually changed, which question remains open and what the next scene must accomplish differently. The next scene's causal plan is rebased on that handoff before prose is written, so chapter 7 follows what the book established, not merely what chapter 1 intended. A scene that contradicts confirmed state is rewritten once, with the contradiction named.

Run a full book from the terminal and inspect every model call

npx vite-node scripts/run-book.ts --writer deepseek-v4.1-flash:cloud \
  --editor mistral-large-3:675b-cloud --chapters 3 --premise "..." --out runs/demo

STAGE design calls=0
STAGE chapter ch1 calls=2
STAGE chapter ch2 calls=15
STAGE chapter ch3 calls=29
STAGE audit calls=46
DONE status=COMPLETE_WITH_WARNINGS calls=47 tokens~362683

Every call is logged with its route, model, prompt size and duration. --out holds manuscript.md, snapshot.json and run.log. --provider gemini runs the same pipeline against the Gemini transport instead, with --writer/--editor as model names.

How it works

Two models, two roles. The writer produces prose. The editor plans, reviews, extracts memory and audits the finished book; without a separate editor the writer judges its own prose, which is the weakest configuration this engine allows. Eight prompts live as files under prompts/, one per stage, and code never hand-builds them — it names a prompt and supplies its variables. Before any prose exists, structural doubts about a scene (no viewpoint, an outcome that changes nothing, a location the memory contradicts) go to a readiness review that turns a blocking verdict into an instruction for the writer instead of stopping the book. Every stage is committed to IndexedDB, so a closed tab or a failed scene costs at most the call in flight: reopening offers to continue, and finished chapters keep their manuscript and memory.

premise → design → plan review → chapter plan → scene → handoff → rebase → next scene → audit → export

Configuration

Set in the app before generation starts.

Environment variables

Requirements

• Node 18+ (required by Vite 6)
• A browser with IndexedDB — the manuscript, its memory and the audit are stored there
• A Gemini API key, or Ollama running locally or in Ollama Cloud
• A model that honours JSON Schema output and returns long prose in your language
• Known to complete a book: deepseek-v4.1-flash:cloud and mistral-large-3:675b-cloud

Limitations

• Business Source License 1.1 — source available, not an OSI open-source licence
• Verified on macOS only; there is no CI, so Linux and Windows are untested
• The construction review can still end a run before a page is written, but only over what code charges — a premise name nobody in the cast answers to; its own remaining objections travel into the contract instead, and the error screen offers a way back to the form
• A model that ignores JSON Schema stalls the run, and a capped output budget fails a call outright rather than returning a short answer
• Prose texture is measured only as repeated phrasing: the audit reports a beat the book returns to at a rate ("breath hitched", seven times), not sentence rhythm or register
• The writer/editor split was only recently made explicit at the call site; the books measured above were written before that fix, with the editor model drafting every scene, so the pairing above is verified to finish a book but not yet as the roles now read
• Editorial gates support revision; they do not guarantee coherence or literary merit

npm test

 Test Files  13 passed (13)
      Tests  141 passed (141)

npm run build && npm run preview