Fabulist

An AI-native writing studio for the desktop, powered under the hood by Claude Code.

Every document is a real Claude Code project folder — versioned with git, rewindable, and shared with an agent that works inside the document's world. You write directly, or you highlight, comment, and ask Claude. Every change Claude proposes is shown to you as a diff that you approve or decline before it touches the text.

Every edit Claude proposes appears in the document as a suggested edit — strikethrough and insertion, accept or decline:

Every save, snapshot, and approved edit is a point you can return to:

Features

• A real editor first — CodeMirror 6 markdown editing with manuscript typography. Your words live in a plain document.md you can open with anything.
• An agent that knows the document — chat in the sidebar; Claude reads, researches, and edits the doc with full project context (each doc has its own CLAUDE.md it loads). Per-document model picker, populated live from whatever your Claude Code version offers.
• Human-approved edits — Claude's proposed changes render in the document itself as Google-Docs-style suggested edits (strikethrough + insertion), with accept/decline (⌘⏎ / esc). Commands and non-document files surface as diff cards in chat. Nothing reaches the document without your explicit approval.
• Comments are the conversation — highlight text and comment, like a shared doc. Claude reads every comment, replies in the thread, and proposes any text change as a suggested edit. Threads are anchored highlights that survive edits and rewrites.
• Versioned and rewindable — autosave checkpoints, named snapshots, and automatic commits for every approved Claude edit. Preview any version as a diff and restore it; restores commit forward, so history is never destroyed.

Requirements

• macOS (Windows/Linux likely work — packaging targets exist — but are untested)
• Node.js 20+
• git on your PATH
• A logged-in Claude Code installation — Fabulist uses your existing claude login and plan; it never asks for an API key

Run it

git clone <this repo> fabulist && cd fabulist
npm install
npm run dev

If your npm config disables postinstall scripts, run node node_modules/electron/install.js once to fetch the Electron binary.

Build a standalone app (.dmg/.zip into dist/):

npm run dist

Your documents live in ~/Documents/Fabulist/, one folder per document:

~/Documents/Fabulist/<document>/
  document.md        ← the document itself
  CLAUDE.md          ← per-document instructions Claude loads automatically
  comments.json      ← anchored comment threads (versioned with the doc)
  .fabulist/         ← app state: agent session id, chat transcript (gitignored)
  .git/              ← full version history

How it works

Layer	Choice	Why
Shell	Electron + electron-vite + React	The Claude Agent SDK is a Node library; running it in the main process is the most direct integration possible — no sidecar, no server.
Editor	CodeMirror 6 (markdown)	Baremetal; its decoration/range-mapping API powers comment highlights that survive edits.
Agent	@anthropic-ai/claude-agent-sdk	The actual Claude Code engine. cwd = the doc folder, resume keeps one continuous session per document, settingSources: ['project'] loads the doc's CLAUDE.md.
Versioning	plain git CLI per doc folder	History = git log, rewind = restore from a rev (always committed forward — restoring never erases history).
Approval gate	SDK canUseTool callback	Read-only tools pass through; file edits and commands surface as inline diff cards the author must approve. comments.json is app-managed and off-limits to the agent.

The approval flow

1. Claude calls Edit/Write on a file → canUseTool fires in the main process.
2. Main computes before/after, sends a permission request over IPC.
3. For the document itself, the editor renders the change inline as a suggested edit (old text struck through, new text inserted) and locks the doc while you review; commands and other files show as diff cards in the chat panel.
4. On accept, the SDK executes the edit; a file watcher picks up the change and the editor updates live; the turn ends with an automatic Claude: <prompt> commit.

Privacy

Everything is local: your documents, comments, history, and chat transcripts stay in ~/Documents/Fabulist/. Conversations with Claude go through your own Claude Code login, exactly as if you'd run claude in that folder yourself.

Development

• npm run typecheck — strict TS over main + renderer
• npm run build — production bundles to out/
• npm run pack — unpacked app build for quick smoke tests (dist/)
• scripts/cdp.mjs — dev harness: launch with npm run dev -- -- --remote-debugging-port=9223, then node scripts/cdp.mjs screenshot /tmp/app.png or node scripts/cdp.mjs eval '<js>' (the renderer exposes window.__store in dev builds)

Source map:

src/main/        Electron main: window, IPC, git, library, comments store,
                 agent.ts (the Claude Code bridge + approval gate)
src/preload/     typed contextBridge API (window.fabulist)
src/shared/      types shared across processes
src/renderer/    React app: Library rail · CodeMirror editor (comment
                 decorations, selection toolbar) · sidebar (Claude chat with
                 approval cards / comment threads / history timeline)

See CONTRIBUTING.md before opening a PR.

License

MIT