The problem
A novel is a large, long-lived project: a hundred thousand words, dozens of characters, facts that change over time, secrets only some characters know, and setups planted two hundred pages before their payoff. Most AI writing tools offer a chat window that forgets, or a "generate chapter" button whose output you can't trace or trust.
So I asked what writing fiction with AI would look like if it worked the way software engineering does. The author stays in charge: AI never silently edits the manuscript, and every suggestion is reviewable and reversible.
Software engineering, applied to fiction
| Software engineering | openAuthor |
|---|---|
| Repository | A project of Markdown scene files that stay useful without the app |
| Commit | A named, restorable revision |
| Pull request and code review | A Change Set, reviewed and applied hunk by hunk |
| Branch and merge | A Storyline, compared with the main manuscript and reconciled |
| Context for an AI agent | Story Memory and a Context Pack the author approves |
| Tests and linters | Continuity diagnostics and editorial passes |
| Rate limits and cost controls | Fail-closed token and cost budgets per run, day, and week |
| Logs and provenance | A record of every AI call's prompt, context, model, tokens, and cost |
What I built
A Python core with no runtime dependencies handles scenes, Story Memory, Change Sets, providers, budgets, retrieval, git operations, and continuity checks. The desktop app is React and TypeScript in a Tauri shell with a Rust bridge. A mobile companion can read, chat, and capture but intentionally can't edit the manuscript, and a small gateway issues short-lived voice sessions so API keys stay server-side.
Story Memory tracks canon facts, who knows what and when, relationships, arcs, and setups and payoffs, each cited to the passages that support it. A manuscript scan proposes entries as suggestions, and nothing becomes canon until the author accepts it.
Status
openAuthor is in active development and hasn't been released. Twenty-five of its twenty-nine tracked capabilities are visible in the desktop app and twenty are covered by automated end-to-end workflows; signed distribution, clean-machine installs, and author pilots are still open.