From the first chapter to the last,
help your AI keep track of the past.

No. This site is an introduction. Drafting, research and revision happen in your AI tool, and the workflow runs through local Python tools. The core scripts work offline; whether your AI service goes online and how it handles your material depends on the tool you choose.

COSMO is free and open source. It has no built-in paid API and never calls a model for you. Your AI tool bills according to its own pricing.

Token usage is hard to predict: the same text tokenizes differently across models, and the amount of context and the number of exchanges both matter. The character budgets in COSMO only limit how much material goes into a reading package; they are not token or cost caps. Try a chapter or two and check your actual usage.

Any agent that can open the whole project folder, read and write files and run commands. Here are some options; pick one you are comfortable with.

Desktop workspaces: WorkBuddy; Claude Code also offers desktop, editor and terminal interfaces.

Agents in an editor: Cursor Agent, GitHub Copilot (VS Code agents) and Cline.

Command-line tools: Codex CLI, Claude Code and Gemini CLI. Cline also has a CLI.

Uploading files to a chat is not enough. For cloud workspaces, confirm that the complete project can be exported or synced back to your machine. If you use another file-enabled Claude mode, check separately that it can read the whole project, run commands and save changes.

All links are official sources, checked on September 23, 2026. The list is not a compatibility certification or a ranking of writing quality. This release is tested on macOS only and requires Python 3.9+; a tool that runs on Windows does not mean COSMO supports native Windows. Try it on a separate copy first.

Not yet. The English website translates the introduction; the system core, templates and user documentation are still primarily in Chinese, and a complete English writing experience is not promised.

The licenses allow you to adapt COSMO to other languages. That means adapting prompts and templates, Chinese paths and field parsing, and the language checks, then validating everything with an example project.

Yes. New books use General mode by default, which checks reader expectations and how each stage develops. It suits relationship, coming-of-age and slice-of-life stories, and it allows a natural ending. Progression and Conflict mode adds checks for opponents, escalation and satisfying payoffs.

Books upgraded from earlier versions are unaffected: older cards keep their original meaning, and books that never enabled this feature do not need to fill anything in.

No. Reviews look at what keeps a reader engaged, and a quiet chapter can move the story through a new feeling, a changed relationship or fresh information. The review pinpoints where a passage loses interest before anything speeds up, and it never piles on conflict or twists mechanically.

The program only flags that a review is due; you and your AI reach the conclusions by reading. Good prose does not need to change just to fill in a form.

No. Research into ranked books is optional; if you skip it, note the reason. If you do it, an AI tool with browsing access picks comparable works, reads legally accessible consecutive passages, records sources, dates and reading scope, and distills general techniques this book could use.

If access is limited, it records what was actually read and does not automatically add a follow-up task to the next cycle. COSMO's own tools never go online and never treat a synopsis as a chapter that was read.

No. The chosen sample enters a candidate first: INIT during book setup, REVISE for a book already in progress. The adopted voice sample changes only after the sample analysis is complete and you approve it.

An existing book that needs its setup redone can explicitly choose INIT. If both kinds of candidate exist, the tool asks you to pick which one this round handles. Voice, genre plugins and story mode are separate choices.

After a chapter commit succeeds, 6.0 moves the candidate into a retained-candidate directory and reports where it is. Anything saved during the commit can be checked there. It never becomes part of the manuscript automatically, and your earlier approval does not cover it.

Regular backups keep the current adopted content, excluding candidates, retained candidates, reading packages and transaction history. Save important unused material separately. A snapshot taken during an interrupted save preserves the scene but does not guarantee a consistent book can be restored from it; check the records before recovering.

Yes. Your agreements and progress are stored in local files, so another AI tool that works with local files can pick up where you left off. There are 4 optional genre plugins: mystery, romance, science fiction and fantasy, and serialized web fiction. You can also use none.

After switching tools, have the new AI read the book's entry files and current state first. Later decisions still follow what you have approved.

No. Start a new book from a saved blank copy or a fresh download, and keep using the existing folder for the book you are writing. Before updating, the system checks the recorded installation baseline; unknown differences or local edits stop the update by default, and recognized book-specific customizations are preserved by rule.

An update never adopts a new story plan for you or changes what you have written.

The 6.0 release self-test passes 446 / 446 checks, covering book setup, reading packages, commits, revisions, review records, synchronization and recovery.

Passing software checks does not prove a novel is good. Characters, prose and reader engagement still need real reading and reader feedback. See the acceptance report (Chinese) and known limitations (Chinese).