开源本地优先的设计工作台，兼容 22 种编程代理与 150+ 设计系统
Open Design 是一款本地优先、开源的代理原生设计工作台，志在成为 Anthropic Claude Design 的开源替代。它不内置代理，而是与本地已有的编程代理（Claude Code、Codex、Cursor、Copilot 等 22 种 CLI）协作，利用 MCP 协议让它们读取设计系统、技能和插件，直接在终端生成原型、仪表盘、演示文稿、图像/视频等制品。支持 BYOK（自带密钥）兼容任意 OpenAI 端点，提供 macOS/Windows 原生桌面应用。内置 100+ 技能、150+ 品牌级 DESIGN.md 系统、261 款插件，适合追求品牌一致性与开发流程可控的设计师与前端工程师。
Open Design (OD) is a local-first, open-source alternative to Anthropic's Claude Design. It turns your laptop into a design studio by integrating with 22 coding-agent CLIs (Claude Code, Codex, Cursor, Copilot, etc.) and shipping 100+ skills, 150 brand-grade design systems, and 261 ready-to-use plugins. Unlike cloud-only tools, OD runs on macOS, Windows, and Linux, with zero telemetry and full privacy.
Open Design（简称 OD）是一个本地优先的开源设计工作台，是 Anthropic Claude Design 的开源替代品。它通过集成 22 个智能体 CLI（包括 Claude Code、Codex、Cursor、Copilot 等），配合 100+ 技能、150 个品牌级设计系统以及 261 个即用插件，把你的笔记本电脑变成一间设计工作室。不同于纯云端工具，OD 运行在 macOS、Windows 和 Linux 上，无遥测、隐私完全可控。
In April 2026, Anthropic released Claude Design — the first LLM that delivered design artifacts directly. It went viral but stayed closed-source, paid-only, cloud-only, locked to Anthropic's model and ecosystem. Open Design breaks this lock: it is Apache-2.0, self-hostable (desktop or Docker), agent-agnostic, and model-agnostic (works with any OpenAI-compatible endpoint via BYOK). The same artifact-first loop — brief, direction, artifact, critique, deliver — now runs on your own hardware with your own agents.
2026 年 4 月，Anthropic 发布了 Claude Design——第一个可以直接交付设计产物的 LLM。它迅速走红，但始终闭源、付费、只限云端、绑定 Anthropic 的模型与生态。Open Design 打破了这种锁定：它使用 Apache-2.0 许可，可自托管（桌面或 Docker），不绑定任何智能体或模型（通过 BYOK 兼容任意 OpenAI 兼容端点）。同样的产物优先循环——需求、方向、产物、评审、交付——现在可以在你自己的硬件上、用你自己的智能体来运行。
OD's core capabilities are organized around three composable concepts: skills, design systems, and plugins. Skills (100+ shipped) define what the agent produces — from web prototypes, decks, and dashboards to images, video, and motion graphics. Design systems (150 brand-grade) are single Markdown files following a 9-section schema (color, typography, spacing, layout, components, motion, voice, brand, anti-patterns) that any render reads. Plugins (261 official) bundle workflows, are MCP-native, and run in any supported agent. Together, they let you switch a brand theme with one click, export to HTML/PDF/PPTX/MP4, and even refresh an existing codebase to match a design spec.
OD 的核心能力围绕三个可组合的概念组织：技能、设计系统和插件。技能（已内置 100+）定义智能体产出的类型——从网页原型、看板、仪表盘到图像、视频和动态图形。设计系统（150 个品牌级）是遵循 9 段架构（色彩、字体、间距、布局、组件、动效、语调、品牌、反模式）的单个 Markdown 文件，每个渲染都会读取。插件（261 个官方）打包工作流，原生支持 MCP，可在任何受支持的智能体中运行。三者组合，一次点击就能切换品牌主题，导出为 HTML / PDF / PPTX / MP4，甚至可以刷新现有代码库使其符合设计规范。
OD runs as a local daemon (Express + SQLite) that exposes an API layer and an MCP stdio server. The frontend (Next.js 16 App Router) communicates with the daemon via /api/* endpoints. When a user sends a brief, the daemon spawns the selected coding-agent CLI (e.g., claude or codex), passing the SKILL.md and DESIGN.md from the current project. The agent streams output back to the daemon, which parses the <artifact> tags and renders them in a sandboxed iframe preview. The BYOK proxy at /api/proxy/{provider}/stream handles model requests with built-in SSRF protection. The desktop app wraps this in an Electron shell with a sidecar protocol for IPC automation.
OD 以一个本地守护进程（Express + SQLite）运行，对外暴露 API 层和一个 MCP stdio 服务器。前端（Next.js 16 App Router）通过 /api/* 端点与守护进程通信。用户提交需求后，守护进程会启动选定的智能体 CLI（例如 claude 或 codex），传入当前项目的 SKILL.md 和 DESIGN.md。智能体将输出流式返回给守护进程，守护进程解析 <artifact> 标签并在沙箱 iframe 预览中渲染。位于 /api/proxy/{provider}/stream 的 BYOK 代理处理模型请求，并内置 SSRF 防护。桌面端通过 Electron shell 封装，并配有 sidecar 协议用于 IPC 自动化。
The fastest way to use OD is the desktop app — download from the website or GitHub Releases, launch, and start typing a brief. For CLI-native workflows, run od mcp install <agent> (e.g., od mcp install claude) to wire the MCP server into your agent. Then inside the agent, you can call od plugin apply od-default --input brief="a landing page" or simply prompt "use open-design to generate a landing page with the Linear design system". For Docker deployment, clone the repo and run docker compose up -d from the deploy/ directory. All approaches share the same skill/plugin/design-system catalog.
最快上手 OD 的方式是使用桌面应用——从官网或 GitHub Releases 下载、启动后即可输入需求。对于 CLI 原生工作流，运行 od mcp install <agent>（例如 od mcp install claude）即可将 MCP 服务器接入你的智能体。之后在智能体内，可以调用 od plugin apply od-default --input brief="a landing page"，或直接提示「使用 open-design 以 Linear 设计系统生成一个着陆页」。Docker 部署则克隆仓库后从 deploy/ 目录运行 docker compose up -d。所有方式共享相同的技能/插件/设计系统目录。
OD excels in agentic design workflows: rapid prototyping, brand-aligned artifact generation, presentation decks, dashboard mockups, and motion graphics. It is also well-suited for teams wanting to migrate design tokens from Figma/Pencil to code, or to refresh an existing codebase to a brand spec. However, note that it is not a replacement for pixel-precision vector editing (Figma's canvas) — it operates at the artifact and code level. The daemon binds to 127.0.0.1 by default; LAN exposure requires explicit configuration. The skill and plugin ecosystem is file-based, so adding custom items is straightforward but requires understanding the SKILL.md and open-design.json formats.
OD 在智能体设计工作流中表现出色：快速原型、品牌对齐的产物生成、演示文稿、仪表盘原型和动态图形。它也非常适合想要将设计令牌从 Figma / Pencil 迁移到代码的团队，或将现有代码库刷新到品牌规范。但请注意，它不能替代像素级精确的矢量编辑（Figma 的画布）——它工作在产物和代码层面。守护进程默认绑定到 127.0.0.1，LAN 暴露需要显式配置。技能和插件生态基于文件，因此添加自定义项很直接，但需要理解 SKILL.md 和 open-design.json 格式。