What it does

This is a self-hosted “AI director” system for writing full novels. You feed it a vague idea; it generates worldbuilding, character rosters, volume outlines, chapter breakdowns, and actual prose through a structured workflow. The stack is React/Vite frontend, Express/Prisma backend, LangChain/LangGraph for agent orchestration, Qdrant for RAG, and SQLite by default (Qdrant optional). There’s also a Windows desktop build if you don’t want to run the monorepo locally.

The interesting bit

Most AI writing tools are fancy autocomplete. This one is closer to a project management system that happens to produce fiction. It has explicit Planner, Tool Registry, Runtime, approval nodes, checkpoint recovery, and state cards. The “writing style engine” is particularly odd: it extracts stylistic features from sample text, compiles them into toggleable rules, and rebinds them to generation pipelines. The README openly admits the target user is “complete novices who don’t understand writing structure,” not seasoned authors — a refreshingly honest product call.

Key highlights

• Auto-director bootstrapping: one-liner inspiration → multiple book-direction candidates → iterative refinement without restarting from scratch
• Style as asset: extract, save, edit, and bind writing styles across projects; not just prompt suffixes
• Structured memory: worldbuilding snapshots, character dynamics with “absence risk,” and knowledge-base feedforward into later chapters
• Model routing: split planning/drafting/review across different providers (OpenAI, DeepSeek, SiliconFlow, xAI)
• Production chain: single main pipeline for chapter execution, batch generation, and quality-debt tracking, with failure reasons exposed

Caveats

• The README notes this is emphatically not a “one-click book machine”; expect to babysit the pipeline
• 95%+ of the README itself was AI-written, which is either transparency or a warning depending on your view
• Windows desktop releases exist but the project is clearly still in rapid iteration (latest release notes are dense with bugfix categories)

Verdict

Worth a look if you’re researching AI-native product architecture, LangGraph workflow design, or the mechanics of long-context consistency. Probably overkill if you just want to experiment with story prompts; likely too rigid if you’re already a disciplined novelist with your own toolchain.