I built a local-first AI writing studio after manually testing how to write a complete book with an LLM
I've been working on Manurio, a local-first tool for creating complete books with LLMs.
It started as an experiment.
I wanted to see if I could actually write a good, coherent book with an LLM — not just generate chapters, but create a story that holds together from beginning to end.
Before building any software, I spent time figuring out two things:
- how authors normally structure and plan a book
- how to break that process down in a way that works well with an LLM
I then went through the whole process manually.
I developed the premise, characters, locations, plot, chapter plans and eventually the actual manuscript, feeding the right information back into the LLM at each stage.
And the result was much better than I expected.
The story actually held together.
That manual workflow eventually became the foundation for Manurio.
Instead of trying to generate a whole book from a single prompt, Manurio takes you through the same kind of structured process:
- book direction and summary
- characters and locations
- master plot
- chapter outlines
- chapter-by-chapter writing
- review and continuity checks
- revisions and version history
- final production/export
You can go through the process yourself and approve the steps along the way, or use an autonomous run where the agents continue working through the book and review their own work.
A big goal for me is also to make it local-first.
Manurio can run in Docker and connect to an OpenAI-compatible endpoint such as LM Studio, so you can create a book using a model running entirely on your own machine. There are instructions on the project page for getting it up and running with Docker if you want to try it yourself.
I've recently been testing it with google/gemma-4-26b-a4b-qat, and I've been getting really good results from it. The book shown in the screenshot was generated with that model through an autonomous run.
The source code is publicly available under the PolyForm Perimeter License 1.0.1. Manurio is free to use and modify under that license, and anything you generate with it belongs to you.
Project page:
https://grunna.gitlab.io/manurio/
Source:
https://gitlab.com/grunna/manurio
This is the first time I'm sharing the project here, and I'd really like to hear what people think.
Especially if you've experimented with local LLMs or long-form generation:
What part of the book-writing process would you want the AI to handle, and what parts would you want to stay in control of yourself?
