Announcing Open Design: An Automation Engine Integrating Local AI with Design Systems
Project Background: An Alternative to Closed-Source Design AI
Anthropic’s Claude Design, released on April 17, 2026, demonstrated the potential of Large Language Models (LLMs) to directly deliver finished designs, garnering widespread attention. However, its closed-source, cloud-dependent, and paid model limits the autonomy of developers and designers. In response, the Open Design project was created. It adopts the same “Artifact-First” philosophy but is built on a fully open, self-hostable, and vendor-agnostic tech stack.
Core Philosophy: Local-First and Artifact-Driven
The core of Open Design is being “Artifact-First,” meaning the entire workflow revolves around the generation and iteration of the final deliverable (e.g., web prototypes, presentations). Unlike the conversational models of traditional AI, it uses structured forms and task planning to capture requirements precisely. Its “local-first” architecture allows the tool to run on the user’s local machine, directly utilizing various coding agents already installed (like Claude Code, Codex, Gemini CLI, etc.). This not only protects data privacy but also allows users to reuse their existing paid model credits.
Technical Architecture: Daemon and Composable Skills
The technical core of Open Design is a daemon process that runs locally. This process starts in the project directory and provides AI agents with real local file system read/write access, command-line (Bash) execution, and network capabilities. This design grants AI agents operational abilities nearly equivalent to a human developer, enabling them to create project structures, read templates, perform code checks, and work within a real disk environment.
The project features a highly modular design, with its capabilities driven by 19 composable “Skills” and 71 brand-level “Design Systems.” The skill library covers a wide range of common scenarios, from website landing pages and dashboards to mobile app prototypes, weekly reports, and OKRs. The design systems include visual guidelines from well-known tech companies like Linear, Stripe, Vercel, and Apple, allowing the AI to directly generate interfaces with high consistency and professional quality.
A Complete Workflow Loop
A typical workflow begins with the user inputting a requirement, such as “create a magazine-style fundraising presentation.” The system first generates an initial form to collect key information. Next, the AI agent selects a visual style, generates a public to-do list, and creates a real project directory on the local disk. Before generating any design drafts, the agent is forced to perform a “pre-flight” check, reading relevant templates and specifications, and reviewing its own output plan from multiple perspectives. Finally, the validated artifact is rendered in a sandboxed mode for the user to preview, save, or export.
Open Source and Deployment
The Open Design project is fully open-source under the Apache-2.0 license, with its code hosted on GitHub (github.com/nexu-io/open-design). It supports multiple deployment methods, including local development via pnpm dev, one-click deployment to the Vercel platform, or running as a single-process production service. This openness and flexibility are intended to encourage community participation and further development, collectively building a powerful AI design ecosystem.
