What it is
Open Design (open-design.ai) is the open-source alternative to Anthropic’s hosted Claude Design product. The repo is nexu-io/open-design, licensed Apache-2.0, current version v0.8.0-preview. Everything runs locally — a desktop app plus a Node daemon — and there is no subscription.
The pitch: you already pay for a coding agent (Claude Code, Codex, Cursor, Gemini, …). Open Design turns that same agent into a design engine. It generates web prototypes, dashboards, mobile screens, decks, magazine posters, sprite animations, motion frames, even videos and audio — written as files to disk that you preview in a sandboxed iframe and export as HTML / PDF / PPTX / ZIP / Markdown.
Architecture — how it works
Open Design has three layers:
- Daemon — a Node 24+ background process. On boot it scans
$PATHfor installed coding-agent CLIs and auto-detects which ones you have (Claude Code, Codex, Cursor, etc.). It loads all 132 skills and 150 design systems and exposes them to whichever agent you point at it. - Skills & systems — every skill is a portable
SKILL.mdfile, every design system is a portableDESIGN.mdfile (9 sections: colour, typography, spacing, layout, components, motion, voice, brand, anti-patterns). Drop-and-restart architecture — add a file, the daemon picks it up. - Desktop / web UI — you type a prompt in the app; the daemon orchestrates the agent; artifacts land in
./.od/projects/<id>/and stream into a sandboxed preview iframe.
The Detect → Discover → Direct → Deliver workflow
- Detect — daemon scans
$PATH, lights up the agents you have, auto-loads 132 skills and 150 design systems. - Discover — turn 1 surfaces a structured question form: surface (web / mobile / deck / poster), audience, tone, scale, brand. No guessing, no “tell me more” loop.
- Direct — turn 2 returns five deterministic visual directions, each with an OKLch palette, a font stack, and a vibe descriptor. Pick one.
- Deliver — the agent writes files to disk. You see a streaming todo card, then a live preview, then export.
Install
Option A — desktop app (recommended for most users). Prebuilt binaries on the GitHub releases page for macOS (Apple Silicon + Intel), Windows x64, and Linux AppImage. Download, install, launch:
# macOS
"/Applications/Open Design.app/Contents/MacOS/Open Design"
# Windows
"C:\Program Files\Open Design\Open Design.exe"
# Linux
./OpenDesign.AppImage
Option B — from source. Requires Node 24+ and pnpm via corepack:
git clone https://github.com/nexu-io/open-design.git
cd open-design
corepack enable
corepack pnpm --version # verify pnpm 10.33.x
pnpm install
pnpm tools-dev run web
Option C — Docker (for headless deploy):
git clone https://github.com/nexu-io/open-design.git
cd open-design/deploy
cp .env.example .env
openssl rand -hex 32 # generate a secure token
# paste the token into OD_API_TOKEN= in .env
docker compose up -d
# now reachable at http://localhost:7456
Usage examples
1. Daemon control from a terminal:
pnpm tools-dev run web # foreground, logs to files
pnpm tools-dev start web # background mode
pnpm tools-dev status # check running processes
pnpm tools-dev logs # tail recent logs
pnpm tools-dev check # diagnostics
pnpm tools-dev stop # stop all sidecars
pnpm tools-dev restart # full restart
# pin ports when running multiple instances
pnpm tools-dev run web --daemon-port 17456 --web-port 17573
2. Generate a SaaS landing page with a specific brand. In the desktop app:
Prompt: A landing page for an AI invoicing tool.
Audience: SMB owners. Tone: trustworthy, modern.
Use the Stripe design system.
Discover (Q form): … surface, audience, tone … → submit
Direct: 5 OKLch palettes proposed → pick “Variant 3 / Editorial”
Deliver: streaming todo card → preview → export HTML
3. Build a deck instead of a page. Same prompt, different skill:
Skill: guizang-ppt (magazine-style, default)
or simple-deck / replit-deck / weekly-update
Prompt: Q2 product review — 12 slides covering shipped features,
funnel, what we learned, what’s next.
Export: PPTX (agent-driven) or PDF (deck-aware print).
4. Bring your own key via the OpenAI-compatible proxy. Useful for DeepSeek, Groq, OpenRouter, self-hosted vLLM — everything goes through:
POST /api/proxy/anthropic/stream
POST /api/proxy/openai/stream
POST /api/proxy/azure/stream
POST /api/proxy/google/stream
POST /api/proxy/ollama/stream
What ships in v0.8.0
- 132 skills across categories — design & marketing prototypes (web-prototype, saas-landing, dashboard, pricing-page, docs-page, mobile-app, mobile-onboarding, gamified-app, email-marketing, social-carousel, magazine-poster, motion-frames, sprite-animation, dating-web, digital-eguide, wireframe-sketch, critique, tweaks), decks (guizang-ppt, simple-deck, replit-deck, weekly-update), ops & office (pm-spec, team-okrs, meeting-notes, kanban-board, eng-runbook, finance-report, invoice, hr-onboarding), plus image / video (Seedance 2.0, HyperFrames) / audio generation.
- 150 portable DESIGN.md systems — Linear, Vercel, Stripe, Airbnb, Notion, Apple, Anthropic, Cursor, Supabase, Figma, Spotify, Tesla, Nike, Shopify, Uber, IBM, NVIDIA, Cohere, Mistral, OpenAI, Google, Meta, X.AI, Raycast, Sanity, PostHog, Sentry, MongoDB and 120+ more, categorised by domain (AI/LLM, Developer Tools, Fintech, E-Commerce, Media, Automotive).
- 16 first-party agent adapters — Claude Code, Codex CLI, Devin for Terminal, Cursor Agent, Gemini CLI, OpenCode, Qwen Code, Qoder CLI, GitHub Copilot CLI, Hermes (ACP), Kimi CLI (ACP), Pi (RPC), Kiro CLI (ACP), Kilo (ACP), Mistral Vibe CLI (ACP), DeepSeek TUI. A BYOK proxy covers any OpenAI-compatible endpoint that isn’t in the list.
- 5 export formats — HTML (inline assets), PDF (browser print, deck-aware), PPTX (agent-driven), ZIP (archiver), Markdown.
- MCP server baked in — wire Open Design into Claude Code, Cursor, or VS Code via Settings → MCP server so other tools can read files from the active project.
First-run layout
The daemon writes a gitignored ./.od/ directory at the repo root. Relocate via the OD_DATA_DIR env var:
.od/
├── app.sqlite # projects, conversations, messages, tabs
├── artifacts/ # timestamped one-off renders
└── projects/<id>/ # per-project working directory (the agent’s cwd)
Media providers & the MCP bridge
Media generation is configured in Settings → Media providers. Defaults include OpenAI gpt-image-2 for image, Seedance 2.0 and HyperFrames for video, plus an audio path. Anything OpenAI-compatible plugs in as a custom endpoint — useful for self-hosted FLUX or local SD setups.
The MCP server (Settings → MCP server) is the other half of the local-first story. Once enabled, Claude Code or Cursor can read files directly from your active Open Design project without re-uploading anything. That means an artifact generated by Codex inside Open Design is immediately readable from Claude Code in another terminal — the source of truth is the file on disk, not a cloud doc.
Why it matters / where I use it
Two reasons this is a category move, not just a clone. One: local-first means your design files, prompts, and brand context stay on your machine. No upload step, no “we may use this for training” clause, no monthly seat. Two: portable DESIGN.md files are the right primitive. A single markdown file describing a system (Linear or Stripe or your own brand) travels between Claude, Cursor, Codex, and Gemini without re-engineering anything. That is the part Figma can’t copy.
I use it for fast brand-mode prototyping — pick a system, hit Discover, get five directions in a minute, ship a static HTML preview to a client the same day. The slide-deck path also replaces about 80% of my Keynote time on internal updates.
Three places it has shipped real output: (1) a SaaS landing-page first draft for a client, generated against the Vercel DESIGN.md system, exported as standalone HTML and handed over in 25 minutes; (2) a 14-slide Q2 product review built with the guizang-ppt skill that I exported as PPTX and only nudged a few headings on; (3) a magazine-poster skill run that generated three editorial layout variants for a single feature article — faster than picking a Figma template and lighter than Photoshop.
Source
- open-design.ai — landing & docs
- github.com/nexu-io/open-design
- Releases — desktop binaries
- open-design.ai/skills/ — full skill catalogue
- open-design.ai/systems/ — all 150 design systems