Fabulist

An AI-native studio for the desktop — make anything with Claude, powered under the hood by Claude Code.

Every project is a real Claude Code project folder — versioned with git, rewindable, and shared with an agent that works inside the project's world. A project can be prose, a website, a slide deck, research, code, data — and a fabulist.json manifest can grow a custom studio around it: typed files, palette actions, rendered panels, live previews. You work 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 lands.

Every edit Claude proposes appears in the file 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 editing for every textual file: Markdown gets a reading page, code gets monospace and syntax highlighting, images get an image view. Your work lives in plain files you can open with anything.
• Studios, not just files — a checked-in fabulist.json turns a project into a studio: typed file kinds with templates, ⌘K palette actions backed by skills, rendered panels (markdown, code, image, live web), and user-run commands with log tabs and a live localhost preview. Built-in starters cover websites, slides, research, essays, novels, and campaigns — or describe your work and the workshop agent designs a studio around it.
• An agent that knows the project — chat in the sidebar; Claude reads, researches, and edits with full project context (each project has its own CLAUDE.md it loads). Per-project model picker, populated live from whatever your Claude Code version offers.
• Human-approved edits — Claude's proposed changes render in the file itself as Google-Docs-style suggested edits (strikethrough + insertion), with accept/decline (⌘⏎ / esc). Commands and other files surface as diff cards in chat. Nothing lands without your explicit approval (unless you trust a studio with auto-edits).
• Comments are the conversation — highlight text and comment, like a shared doc. Claude reads every comment, replies in the thread, and proposes any 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 projects live in ~/Documents/Fabulist/, one folder per project:

~/Documents/Fabulist/<project>/
  *.md, *.html, …    ← the work itself, any file kind, nested folders fine
  fabulist.json      ← optional studio manifest (file types, actions, panels…)
  CLAUDE.md          ← per-project instructions Claude loads automatically
  .claude/skills/    ← reusable skills the studio's actions invoke
  comments.json      ← anchored comment threads (versioned with the project)
  .fabulist/         ← app state: agent session ids, chat transcripts (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 project folder, resume keeps continuous sessions per thread, settingSources: ['project'] loads the project's CLAUDE.md.
Versioning	plain git CLI per project 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 user 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 focused file itself, the editor renders the change inline as a suggested edit (old text struck through, new text inserted) and locks the file 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 projects, 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