Open-source design editor built on one premise: a design is a node tree, not a picture.
Design tools treat the canvas as the product and the file as an implementation detail — a proprietary binary only the vendor's software can fully read, behind APIs that fight automation. OpenPencil inverts that. The document is a node tree you can open, query, transform, lint, and diff; the editor is one client of it. A designer moves things on the canvas; a script renames every color token; an AI agent builds a screen — all through the same small set of operations, and each change is data you can review like a code change.
It opens Figma's .fig files natively, runs as a ~7 MB desktop app or a
browser tab, and never sends your files anywhere. The editor, the engine,
the file codec, the CLI, the MCP server, and the Vue SDK are all MIT.
brew install open-pencil/tap/open-pencilOr use the web app — no install, no account.
Try it online · Download · Documentation · Roadmap
| Project | What it is |
|---|---|
| open-pencil | The editor — desktop and web app, plus the engine, CLI, MCP server, and Vue SDK as workspace packages |
The same repository publishes the programmable surface to npm:
| Package | What it does |
|---|---|
| @open-pencil/cli | Headless CLI — inspect, query (XPath), lint, analyze, export, and convert .fig/.pen files; control the running editor over RPC |
| @open-pencil/mcp | MCP server (stdio + HTTP) — 100+ design tools for Claude Code, Cursor, Windsurf, and any MCP client |
| @open-pencil/core | The engine — scene graph, renderer, layout, file formats, tools |
| @open-pencil/vue | Headless Vue SDK for embedding OpenPencil or building custom editors |
| Project | What it does |
|---|---|
| skills | Agent skills — teach AI coding agents to inspect, modify, and export design files |
| homebrew-tap | Homebrew tap for the desktop app |
| twirlwind | Tailwind v4-first CSS-to-utility-class serializer, used by JSX/Tailwind export |
The rendering and layout stack OpenPencil runs on, maintained here:
| Project | What it does |
|---|---|
| canvaskit-webgpu | CanvasKit WASM build with Skia Graphite + Dawn (WebGPU) backend |
| skia | Skia fork carrying the Graphite/Dawn CanvasKit patches |
| yoga | Yoga fork with CSS Grid support, powering auto layout |
.fig / .pen document (node tree)
├── editor — desktop (Tauri) and web app, real-time P2P collaboration
├── @open-pencil/core — scene graph, Skia renderer, Yoga layout, codecs, 18-rule linter
├── @open-pencil/cli — tree / query / lint / analyze / export / eval from the terminal
├── @open-pencil/mcp — the same operations as MCP tools for AI agents
└── @open-pencil/vue — headless components for building your own editor
Every surface — canvas, CLI, MCP, SDK — drives the same operations on the same tree. That is the whole design: nothing the editor can do is closed to a script, and nothing a script does bypasses what the editor checks.
OpenPencil is also the design layer of a larger open stack — a web stack built so that software written by AI (and by people) can be checked, not just generated. The argument and the map live in Building Blocks for the Future Web.
MIT across the organization. Forks (skia, yoga) keep their upstream
licenses.
