Open Design: A local-first workspace that turns coding agents into a design studio

What problem Open Design solves

Open Design targets a specific gap: coding agents are good at writing code, but they usually have no structured way to turn a brief into a designed artifact. The README positions it as an open-source alternative to Claude Design, where the agent loop is discover the brief, lock the direction, stream the artifact, critique, deliver. Instead of pushing pixels on a canvas like Figma, Open Design treats the design as a filesystem of functional skills, rendering templates, design systems, and plugins. The intended user is a developer or designer who already works with a CLI-based agent and wants that agent to produce prototypes, landing pages, dashboards, slides, images, and video exports. It is local-first, meaning the workspace runs on your laptop, and the generated files stay in your working directory. This is not a replacement for Figma in the sense of collaborative editing; it is a different model where the agent writes real CSS and components, and you review the rendered output in a sandboxed iframe.

How the agent-native loop works

The core mechanism is that Open Design does not generate designs itself. It provides a workspace where an external coding agent, such as Claude Code or Codex, does the generation. The README describes two integration paths: skills, CLI, and MCP for agents that consume Open Design, and native runtime adapters for agents that Open Design launches directly. DeepSeek Harness is a first-class native runtime through the official dsh CLI, with structured streaming, model discovery, cancellation, and session resume. The workspace has a Studio view where the conversation, generated files, and live preview stay together. There are six artifact types: prototype, deck, mobile app, image, document, and HyperFrame. Each type has a specific preview mode, such as a device preview for mobile apps or a thumbnail view for decks. The generated files are real HTML, CSS, and components, not raster images. Exports go to HTML, PDF, PPTX, and MP4. The design system is defined in a DESIGN.md file, which acts as a brand contract that the agent reads and follows. This is a significant departure from traditional design tools: the source of truth is a text file, not a binary project file.

Getting started with the CLI and MCP

The README gives concrete commands for connecting Open Design to coding agents. For Claude Code, you run `od mcp install claude`. For Claude Desktop, the command is `od mcp install claude-desktop`. The `od` CLI is the entry point, and it installs the Model Context Protocol server for the target agent. The README also lists a table of supported agents, including Codex CLI, Cursor, OpenCode, Qwen, Copilot, and others, but the truncated portion cuts off before showing the full setup commands for each. The pattern is clear: you install an MCP server so the agent can call Open Design tools. For agents that Open Design launches directly, there are native runtime adapters, with DeepSeek Harness as the first-class example. The Home page lets you choose an artifact type, enter a brief, and set the design system, working directory, and model before you start. There is also a Plugins page where you can browse official skills and launch a workflow with a Try it button. The setup is command-line driven, which fits the target audience of developers who are comfortable with a terminal.

The export pipeline and production readiness

Recent release notes emphasize export reliability. Version 0.20.2 is titled Production Exports, Under a Minute, and version 0.21.0 is Reliable, Start to Finish. This suggests that export speed and consistency have been a focus. The README lists HTML, PDF, PPTX, and MP4 as export formats. The MP4 export comes from HyperFrames, which are code-driven motion graphics. That means video is not rendered in a traditional timeline; it is generated from code that animates. The export pipeline is a key differentiator because you get a deliverable file, not just a preview. However, the README does not specify the resolution, frame rate, or codec for MP4 exports. It also does not state whether PDF exports preserve vector graphics or rasterize them. The claim of under a minute for production exports is a marketing line from the release title, and the README does not provide benchmark numbers. You should test export quality with your own content before relying on it for client work.

Limitations and cases where it is the wrong tool

The most obvious limitation is that Open Design is single-page by design. The README says it delivers single-page artifacts, which means multi-page documents or complex web apps with routing require the agent to manage state within one HTML file. That is a constraint for large dashboards or full websites. Another limitation is the dependency on external model services. The app itself is open source, but the README heavily promotes OpenDesign Cloud, a paid service for agent and image models. You can bring your own key (BYOK) for OpenAI-compatible endpoints, but the default experience is tied to their cloud. If you need fully offline operation, you must configure a local model that supports the same tool-calling protocol, which is not documented in the README. Also, the collaborative aspect is missing. There is no mention of real-time multi-user editing or version control integration beyond the local filesystem. If your team expects Figma-like collaboration, Open Design will not provide it. Finally, the artifact types are fixed: prototype, deck, mobile app, image, document, and HyperFrame. There is no generic canvas or free-form layout mode, so if you need to design a complex vector illustration or a data-heavy report with precise layout control, the agent's output may not meet your standards.

Alternative: Claude Design and the agent-native approach

The README explicitly positions Open Design as an open-source alternative to Claude Design. Claude Design is Anthropic's closed implementation of the same agent-native loop, but it is not available as a standalone product you can install. The difference is that Claude Design is tied to Anthropic's models and platform, while Open Design is agent-agnostic. It supports Claude Code, Codex, Cursor, OpenCode, Qwen, Copilot, and many others, according to the compatibility table. That is a real difference: you are not locked into one vendor. Another alternative is Figma, which is a visual canvas tool. Figma gives you pixel-level control and real-time collaboration, but it does not generate code from a brief. Open Design flips that model: the agent writes code, and you review the rendered output. If your workflow is code-first, Open Design fits. If your workflow is visual-first, Figma or a similar canvas tool is more appropriate. The choice is not about which is better; it is about where the source of truth lives. Open Design puts it in a text file and a set of skills; Figma puts it in a binary document.

Maintenance, license, and upgrade cost

The project is under Apache-2.0, which is a permissive license. That means you can use, modify, and distribute the code, including for commercial purposes, as long as you preserve the license notice. The repository is active, with the last push on 2026-08-25 and a release cadence of about four days between 0.20.1 and 0.21.0. That suggests a fast-moving project, which has implications for upgrade cost: you will need to track releases to get fixes and new features. The release notes mention new model integrations and export improvements, so the upgrade path is not just bug fixes but also feature additions. However, the README does not document a migration process or a changelog beyond the release titles. The app is a native desktop application for macOS and Windows, so updates are likely delivered through the app or the CLI. There is no mention of auto-update or package managers like Homebrew or winget. You should plan to manually download new versions from the releases page. The dependency on external model services means your operating cost is tied to the model pricing, which is separate from the open-source code. The README mentions a two-week unlimited access promotion for DeepSeek models, but that is a temporary offer, not a permanent free tier.

Editorial conclusion

Adopt Open Design if you already rely on a coding agent like Claude Code or Codex and want those agents to produce design artifacts without switching to a canvas tool. Skip it if your team needs collaborative, multi-user editing or pixel-level control, because the artifact model is single-page and code-driven. Before adopting, verify that your preferred agent is in the supported list, confirm the export formats cover your delivery needs, and check the license terms for any commercial use of generated assets. The project is Apache-2.0, so the code itself is permissive, but the model services and cloud features are separate paid offerings. Start with a small prototype to see if the agent-native loop matches how your team works.

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes