Narcooo/inkos

InkOS is for writers who have outgrown pasting a chapter request into a chat window and want a repeatable project pipeline for novels, short stories, scripts or interactive fiction. It is a substantial system with a migration path from 1.x, so it rewards a careful trial rather than a casual install.

The problem

A writer working on a serialized novel wants each new chapter to stay consistent with earlier chapters, character decisions and plot threads. Chasing that by hand means re-reading old material, manually restating rules in every prompt, and separately asking for a review before revising. InkOS moves those steps into a workspace: you create a book with a brief, the system generates world, character and volume plans, then chapters are written, reviewed, revised and settled as separate recorded actions. The result is a draft plus the planning and review artifacts around it, rather than a single chat transcript.

What's different

InkOS exposes writing, review, revision and state settlement as individually callable, traceable actions on one shared harness, and organizes context into protected and compressible layers so longer books do not lose their footing. A multi-line plot mode can generate two to five isolated future branches before the next chapter, letting you compare beats, character decisions and risks side by side; adopting a branch only writes a candidate plan file and does not touch the manuscript or canon project instructions. Short stories, covers, interactive worlds and translation are reachable from the same Studio/CLI surface, and heavy actions require confirmation before writing.

The trade-off is configuration weight. You must choose a provider and model, and the project checks that a model is compatible with its service instead of silently sending requests to the wrong endpoint. Skills added as SKILL.md files only provide instructions and reference material; they do not gain execution rights or bypass the confirmation gate. That design is sensible, but it means the product is a harness you configure, not a chat app you just start using.

Who should try it

Good fit: a writer producing serialized Chinese fiction or a scripted interactive story who wants planning, review and revision steps recorded and revisable across a long project. The branch-comparison mode is especially relevant if you routinely make structural choices before drafting. Poor fit: someone who wants a lightweight one-off writing prompt, or a non-technical user expecting a hosted web app. The README points to a web demo for a quick look, but the documented workflow centers on a local CLI/Studio setup web version.

A second option changes the audience: InkOS is published as an OpenClaw skill and can be called from agents compatible with that skill interface, so a Claude Code or OpenClaw user can route requests through the shared inkos interact entry point rather than learning the CLI directly ClawHub listing.

Requirements and limitations

Node.js 22.16 or newer is required for the npm install path npm package. Desktop users need a capable local environment, and every generation consumes a configured LLM service; the README's examples reference DeepSeek models and covers via a paid image service, so API cost and access are part of the setup rather than an afterthought. A local model is supported only through compatible OpenAI, Anthropic or custom endpoints, with protocol and streaming tested at configuration time; a server that lacks structured output or cuts streaming will surface explicit errors instead of silently producing partial text.

Upgrading from 1.x is a separate task: run a dry-run migration listing, then apply, with backups of original directories; migrated projects missing runtime state enter as drafts and need state rebuilding before production. The license is AGPL-3.0, which matters if you plan to build a service around the code license.

HotRepo's take

The only dated HotRepo observation is 10,832 stars on the 2026-10-26 daily trending board, +107 stars that day; there is not comparable daily history to judge whether interest is rising or fading. That snapshot alone supports the project's existence and a single moment of visibility, not a maturity claim. Before committing a long manuscript, verify one thing the README does not settle: how well the protected/compressible context layer holds up for your book's length, by testing the review-and-revise loop on a few chapters.