Three things worth knowing
- Open Design is an open-source, local-first alternative to Claude Design, now at 57.4K stars and 6.5K forks, just eight weeks after its first commit.
- It ships 259+ skills, 142+ design systems, and auto-detects 16 AI coding CLI tools on your
PATHwith zero configuration. - If your team wants LLM-driven design artifact generation without cloud lock-in or provider dependency, this is the most complete community answer I've seen.
Anthropic released Claude Design in April 2026. By June, a community project had built an open-source version of it with 310 contributors and 57,400 GitHub stars.
57K stars in eight weeks on a design tool is a real signal. Claude Design showed what LLM-driven design artifact generation looks like in practice: prompts producing real decks, prototypes, and dashboards. What it didn't offer was self-hosting, model flexibility, or data residency control. Cloud-only, locked to Anthropic's models, no API key substitution.
Open Design is the community's answer.

What Happened
Open Design reached v0.9.0 on June 2, 2026, the 10th release since the project launched in April 2026. It now has 1,837 commits from 310 contributors across 6,500 forks.
The tool runs via pnpm tools-dev, deploys its web layer to Vercel, and stores everything in a local SQLite database. A local daemon scans your PATH on startup, detects available coding agents, and wires them into a skill-driven design workflow without any manual configuration. It's also available as a prebuilt desktop app for macOS (Apple Silicon) and Windows (x64) via open-design.ai, and runs via Docker for teams that want it without installing Node or pnpm.
What I'd flag: the project self-describes as early, with comment-mode surgical edits and a tweaks panel still in progress. The star count is real, but the maturity level is v0.9, not v1.
Key Features
- 16 auto-detected coding agents: Claude Code, Codex CLI, Cursor Agent, Gemini CLI, OpenCode, Devin for Terminal, Kiro CLI, and nine others. The daemon finds them on your
PATHwith zero configuration. Swap between agents via the model picker without changing skills or design systems. - 259+ skills as plain files: Each skill is a
SKILL.mdfolder dropped into theskills/directory. The core library covers prototypes, decks, mobile apps, dashboards, email templates, finance reports, and more. Thehtml-pptskill family adds 15 full-deck templates, 36 themes, 31 single-page layouts, and animated motion graphics. A new team member drops a folder intoskills/, restarts the daemon, and the skill appears in the picker. - 142+ design systems in Markdown: Product-grade systems for Linear, Stripe, Vercel, Airbnb, Tesla, Notion, Apple, Anthropic, Cursor, and more, each defined as a single
DESIGN.mdfile. - Media generation pipeline: Image (
gpt-image-2), video (Seedance 2.0, HyperFrames HTML-to-MP4), and audio surfaces wired into the same chat interface. Outputs land as real files in the project workspace. - Claude Design ZIP import: Drop a claude.ai export onto the welcome dialog. The daemon extracts it into a real project so your local agent can continue editing.
- BYOK proxy with SSRF protection: Multi-provider API proxy supports Anthropic, OpenAI-compatible, Azure OpenAI, Google Gemini, and Ollama endpoints. Loopback is allowed for local LLM providers; non-loopback private ranges are blocked.
Why It Matters
Open Design treats the coding agent as a replaceable component. A team using Claude Code today can switch to Codex or Cursor Agent with one click, without touching their skills or design systems. Design artifacts never leave the developer's machine unless explicitly exported. For teams with data residency requirements, those two things matter more than any feature comparison.
The skill library as version-controlled Markdown files is the part I find most practically useful. Prompt engineering usually lives in someone's head or a shared doc no one maintains. Open Design puts it in a folder with a defined format, which means it can be reviewed, versioned, and shared across a team like any other code.
Example Use Case
A frontend team needs a seed-round pitch deck. They open Open Design, select the guizang-ppt deck skill, and pick the Stripe design system. The turn-1 question form locks in surface, audience, and tone before the model writes anything. The agent picks a visual direction from five deterministic OKLch palettes, streams a live todo plan into the UI, and emits a single-file HTML deck with magazine-style layouts.
The result exports as HTML, PDF, or PPTX. The whole flow uses Claude Code locally, costs only API tokens, and the project persists in SQLite for tomorrow's edits.
This is the workflow I'd walk through with a team asking why they'd use this over a SaaS design tool. The output is the same quality. The data stays local. The cost is API tokens.
Competitive Context
Claude Design is cloud-only and locked to Anthropic's Opus 4.7 model. For teams that need to self-host, use their own API keys, or swap models, Open Design is currently the only local-first option with this level of skill and design system coverage.
Cursor operates as an IDE with AI assistance built in. Open Design treats Cursor Agent as one of many possible backends for a design-specific workflow, auto-detecting cursor-agent on PATH and dispatching prompts through it.
OpenCoDesign, the other open-source Claude Design alternative, ships as an Electron app bundling a specific AI runtime. Open Design avoids bundling an agent entirely, delegating to the user's existing CLI installation. A team running multiple agents gets the same skill library working across all of them.
My Take
310 contributors and 57.4K stars in eight weeks tell me this project landed on a real problem. The v0.9 label is honest: some features are still in progress. For teams using Claude Code, Codex, or Cursor that want design artifact generation without cloud dependency, evaluate it now.
I'm curious whether the skill-as-file model actually changes how teams share design workflows, or whether it just shifts the maintenance burden from docs to directories. Worth watching.
Open Design handles the artifact. Cosmos handles the system.
Free tier available · VS Code extension · Takes 2 minutes
Written by

Paula Hingel
Technical Writer
Paula writes about the patterns that make AI coding agents actually work — spec-driven development, multi-agent orchestration, and the context engineering layer most teams skip. Her guides draw on real build examples and focus on what changes when you move from a single AI assistant to a full agentic codebase.