InkOS is an AI agent system for story writing and long-form translation. Serialized novels, short stories, scripts, storyboards, fan fiction, style imitation, continuations of existing works, interactive films, and open-world interactive fiction all start from one workspace, which you drive through a web Studio, a terminal UI, or a CLI. It writes in Chinese by default and also writes natively in English.
Version 2.0 runs every work type on one production harness built around the Pi agent loop. The model interprets intent, proposes actions, and calls tools, while InkOS handles confirmation, context assembly, state, atomic writes to disk, and checks that the output actually exists. Each work type keeps its own methods as standard SKILL.md skills (19 are bundled), and a project skill with the same ID overrides a built-in one.
A long novel is produced chapter by chapter through a chain of agents:
| Agent | Role |
|---|---|
| Radar | scans platform trends and reader preferences (optional) |
| Architect | creates the story frame, rules, characters, and long-term control files |
| Planner | reads author intent, current focus, and retrieved memory to set the chapter's goals |
| Composer | selects context from state, control documents, and materials, and records a trace |
| Writer | drafts the chapter from the composed context |
| Settler | submits evidence-backed state changes through a typed tool |
| Reviewer | returns evidence-based qualitative observations against intent, canon, and state |
| Reviser | produces a new version against an explicit revision goal |
Each book keeps authoritative structured state (current state, foreshadowing hooks, chapter summaries) validated with Zod schemas, a human-readable Markdown projection, and a rebuildable SQLite FTS5 and BM25 retrieval index. A continuity check flags a character recalling events they never saw, or using an item lost two chapters earlier. Chapter text and state are committed together as one atomic file set, so a failure never leaves state advanced without the matching text.
InkOS is also published as an OpenClaw skill that Claude Code, OpenClaw, and other compatible agents can call, and that entry point uses the same execution core as the terminal UI and Studio.
Features
- Studio Chat: one action surface for chat, book creation, short stories, covers, and interactive worlds, with confirmation before heavy actions and previews of generated output
- InkOS Play: open-world and branching interactive fiction with world state, characters, items, evidence, relationships, time, HUD, and generated illustrations, defined in natural language
- Branch exploration: generates two to five isolated future branches before the next chapter and compares their beats, decisions, risks, and fit with author intent
- InkOS Short: produces a complete short story with outline, review notes, blurb, cover prompt, and an optional cover image
- Continuity review and de-AI pass: records observations on character memory, item continuity, foreshadowing, pacing, and emotional arcs, and flags AI-style phrasing for revision
- Style imitation: compiles reference texts into an evidence-backed style guide and binds it to a book
- Author control documents: editable long-term intent and current-focus documents per book, compiled into each chapter's goals and context
- Word-count tracking: target word counts with an allowed range, counted in Chinese characters or English words, without truncating text
- Import and fan fiction: imports existing chapters and rebuilds state, summaries, and character relationships, and starts fan fiction in canon, AU, OOC, or CP modes
- Per-agent model routing: each agent can use a different model and provider, with fallback to the global model
- Daemon and notifications: a background loop writes chapters, pauses on issues that need a human, and pushes notifications to Telegram, Feishu, WeCom, or signed webhooks
- Provider bank: built-in settings for Gemini, Moonshot, MiniMax, Zhipu, DeepSeek, OpenRouter, Ollama, and others, plus custom OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages endpoints
