Your vector editor now takes pull requests from Claude

What it does

OpenPencil is an AI-native vector design tool built around an infinite canvas. You describe a UI in plain text, and the tool generates it in real time while multiple AI agents work on separate sections—hero, features, footer—simultaneously. The resulting .op files are JSON-based design-as-code artifacts that export to React, Vue, Svelte, Flutter, SwiftUI, and others.

The interesting bit

The orchestrator decomposes complex pages into spatial sub-tasks and dispatches them to concurrent “Agent Teams,” each with their own canvas indicator. It also adapts its prompts to the model tier: Claude gets full reasoning prompts, GPT-4o and Gemini get thinking disabled, and smaller models like Qwen or Llama receive simplified nested-JSON instructions so they don’t hallucinate the layout.

Key highlights

• Built-in MCP server (pen-mcp) exposes the canvas to Claude Code, Codex, Gemini CLI, and Copilot, so you can read and mutate .op files from your terminal
• .op files are diffable JSON with CSS custom properties; the CLI (op) supports piping design DSL from stdin
• Multi-platform code export covers web frameworks, mobile (React Native, Jetpack Compose, SwiftUI), and desktop
• Includes an embeddable headless engine (pen-engine) and React SDK (pen-react) if you want the canvas inside your own app
• “Anti-slop” cross-generation diversity tracking to keep the agents from repeating the same boring layout

Caveats

• The README notes a namesake project also called OpenPencil that focuses on Figma-compatible real-time collaboration; this one is strictly design-to-code
• Auto-layout and some canvas features are mentioned but the feature list is truncated, so the full scope of vector editing capabilities is unclear
• Desktop builds rely on Electron, which may not excite everyone

Verdict

Worth a look if you want a generative design tool that treats UI as version-controlled code and doesn’t lock you into a single model provider. Skip it if you’re after a mature, Figma-like collaborative pixel editor—this is closer to an AI-native IDE for visual interfaces.