Scriptorium

A Windows desktop app where AI agents run a small publishing house: they pick topics, research, outline, write, review, rewrite and publish books, gradually filling a library across every genre, fiction and non-fiction, children's books included. You watch them at work in a pixel-art office and click into any agent to see its full live log.

Status: MVP built on branch SCP-02_mvp-text-only (not committed yet). Phases 0 to 9 are done and verified: app scaffold, data folder, model layer with budget caps, the job queue with agents and a scheduler, the book pipeline from bedtime stories to novels with book memory, the pixel-art office and library, the on-demand researcher (Wikipedia and Tavily search, notes with sources), topics in full, the non-fiction variant (fact-checker, references), and code-drawn cover patterns and chapter ornaments. Real books were made on local LM Studio models for 0 USD. Packaging (phase 10) works (npm run dist), but an overnight run, reboot resume and cloud models are not verified yet. See .claude/tickets/SCP-02-scriptorium-mvp-milestone-1-text-only/ (plan, changelog, handoff).

• Design and every decision so far: docs/design.md
• Current handoff for the next session: .claude/tickets/scriptorium-mvp/handoff.md

In short

• Models: cloud APIs (DeepSeek, OpenAI, Anthropic, Gemini, Mistral, Kimi, Qwen, GLM, OpenRouter) and local ones (LM Studio, Ollama). A default model per agent type, an override per agent.
• Agents: editor-in-chief, idea generator, architect, writers, an on-demand researcher, continuity checker, fact-checker, editors, child-safety and read-aloud reviewers, publisher. Many work in parallel.
• Shared state: plain markdown files on disk. Agents, jobs, bibles, research, reviews and logs are all files.
• Books: up to 150k words, written scene by scene from a book memory (outline, story bible, summaries, RAG, research notes). Output is EPUB with a template cover.
• Budget: a spend counter, caps (40 USD a month, 5 USD a day), and pause and stop buttons.
• Library: inside the app for now. Read, open the folder, send to Kindle, rate. A public side comes later.
• Stack: Electron, TypeScript, Phaser 3, markdown files, and SQLite with sqlite-vec as a rebuildable search index.

Development

Windows 11. Node is only available through fnm (v24). In PowerShell, in every new shell:

fnm env --use-on-cd | Out-String | Invoke-Expression
npm install

• npm run dev: builds the engine, then opens the Electron app. The window shows the engine heartbeat.
• npm run engine -- --data <dir>: runs the engine headless with plain Node, no Electron. It builds out/engine/index.js first, creates the data folder from seed/ (existing files are never overwritten), and keeps running with a heartbeat on stdout. Add --once to create the folder and exit. Press Ctrl+C to stop.
• npm run engine -- probe [--model provider/model]: streams one short reply from LM Studio (and DeepSeek when a key is set), embeds one string, and writes a spend row.
• npm run engine -- reindex --data <dir>: rebuilds the search index (index/library.sqlite, a cache) from the markdown files, using the embedding model. npm run engine -- rebuild-reader --data <dir>: writes the in-app reader folder (books/<slug>/out/reader/) for books published earlier.
• npm run key:set <ENV_NAME>: stores an API key in the Windows credential store (hidden prompt). The engine looks at the environment variable first, then the store.
• To make a book: switch a topic on (active: yes in topics.md in the data folder, or the tick box in the app's Topics panel), add your own topic with "Add a topic" in that panel (it goes into a ## My topics table in topics.md), or drop an idea file into ideas/, then run the engine. Formats: bedtime story (toddler and 3-6), short story, chapter book, and the long ones (middle grade, novella, novel) that use the book memory. Finished books are in books/<slug>/out/<slug>.epub.
• npm test, npm run typecheck, npm run lint, npm run build.

Data folder: --data <dir>, else the SCRIPTORIUM_DATA environment variable, else ~/Scriptorium.

If npm run dev says "Electron uninstall", the Electron binary wasn't downloaded: run node node_modules/electron/install.js.

The SQLite spike (node:sqlite + sqlite-vec) can be rerun with node scripts/spike-sqlite.mjs and npx electron scripts/spike-electron.cjs.