What it does OpenPencil is an open-source vector design tool built around AI generation rather than manual drawing. You describe a UI in plain text, watch it stream onto an infinite canvas, then tweak by chatting with the selection. It exports to React + Tailwind, Vue, Svelte, Flutter, SwiftUI, Jetpack Compose, or React Native. There’s also a CLI (op) that accepts piped input, an MCP server for terminal-based agents, and an embeddable SDK (pen-engine / pen-react) if you want the renderer in your own app.

The interesting bit The “Agent Teams” feature splits a complex page into spatial sub-tasks—hero, features, footer—and dispatches multiple AI agents to work in parallel, each with its own canvas indicator. The orchestrator also adapts prompts per model capability: Claude gets full reasoning prompts, GPT-4o and Gemini lose the thinking tags, and smaller models (Qwen, Llama, MiniMax) receive simplified nested-JSON instructions. It’s a pragmatic admission that not all LLMs are created equal.

Key highlights

• .op files are plain JSON—diffable, Git-friendly, and generate CSS custom properties on export
• One-click MCP server install into Claude Code, Codex, Gemini CLI, OpenCode, Kiro, or Copilot CLI
• Docker variants bundling specific AI CLIs (Claude, Codex, Copilot, etc.), up to ~1 GB for the full kitchen sink
• Desktop builds via Electron for macOS, Windows, Linux; web app runs on Bun
• 15-language UI localization, which is unusually thorough for a 3k-star project

Caveats

• The README notes a name collision: another “OpenPencil” project exists with Figma-compatible real-time collaboration, so check which repo you’re actually starring
• Several Docker image sizes are listed as “—” rather than actual numbers, and the agent-native component optionally requires Zig 0.14 to build from source
• The canvas feature list trails off mid-sentence in the provided README (“Auto-layout — vertical/horizontal with gap,”), suggesting the docs are still being written

Verdict Worth a look if you treat design as a generative step in your build pipeline rather than a handcrafted artifact. Skip it if you need mature real-time multiplayer editing or deeply precise vector tooling—this is AI-native first, traditional-design second.