Writing with sources, on a local AI.
A thesis, a nonfiction book, a long report: the writing is only half the work. The other half is not losing track of what your sources actually say.
Long writing projects fail in the middle, and they fail at the sources. Chapter four needs the thing you read in March, and nobody remembers which paper it was. A local research notebook turns that problem into a question: ask your library, get the passage back with its citation, and write from it while it is on the screen. What the tool will not do is write the chapter. It grounds your drafting in your own evidence, and the draft, the argument and the voice stay yours.
The middle of a long project.
The start of a writing project is fun: reading, collecting, outlining. The end is editing. The middle is where it gets lost, because the middle is where you have read more than you can hold and written less than you need. The stack of sources has become a fog, and the draft stalls on a fact you cannot place.
What helps is not a better writing model. It is keeping the library searchable and next to the draft, so that "where did I read this" costs a question instead of an afternoon. That is a retrieval problem, and it is the part a grounded tool genuinely does well.
How the loop fits the writing desk.
Everything you have collected lives in one local library: the papers, the reports, the transcripts, the scanned pages. While drafting, you ask it questions. The answers cite the passages they drew on, so a quote that goes into the manuscript arrives with its source already attached, and checking it is a click rather than a search session.
Notes sit in the same window as the sources they came from. That is where the writing actually happens, in the exchange between reading and drafting, and it is why Istor is one window rather than a chat app beside an editor. When you need what the open web has, you trigger a fetch yourself and it becomes a source like any other, cited and kept.
And the draft itself is private by construction. Unpublished work is the most exposed file a writer has, and on a local tool it never leaves the machine. Neither does the library it grew from.
What the tool does not do.
It does not write the argument. A chapter is a claim about what the evidence means, and a model asked to generate one produces prose shaped like an argument, verified by no one. What it does instead is refuse that job quietly: it answers from your sources or it says the answer is not there, and every claim it does make points back to the passage it came from. That is the difference between a tool that writes for you and a tool you can write with.
The model on your GPU is also not a frontier model. Istor's roster is sized for a GTX 1650 class card, and the trade is deliberate: a good local model you can verify beats a distant one you cannot, because in a writing project every borrowed sentence has to survive a citation check anyway.