Open Design — Local-First Open-Source Alternative to Claude Design (nexu-io)

Source: Open Design GitHub README (nexu-io/open-design) Repo: github.com/nexu-io/open-design Stars: ~24.3k (per GitHub at ingest, 2026-05-04) License: Apache-2.0 Language: TypeScript

A local-first, open-source alternative to Anthropic’s Claude Design from the nexu-io organization. Open Design (OD) ships 31 composable Skills and 72 brand-grade Design Systems on top of any of 15 coding-agent CLIs already on the user’s PATH — Claude Code, Codex, Devin, Gemini CLI, OpenCode, Cursor Agent, Qwen, GitHub Copilot CLI, Hermes, Kimi, Kiro, Kilo, Mistral Vibe, DeepSeek, Pi. Same artifact-first loop as Claude Design, none of the closed-source / cloud-only / single-vendor lock-in.

What it is

The README’s positioning thesis (verbatim):

Anthropic’s Claude Design (released 2026-04-17, Opus 4.7) showed what happens when an LLM stops writing prose and starts shipping design artifacts. It went viral — and stayed closed-source, paid-only, cloud-only, locked to Anthropic’s model and Anthropic’s skills.

Open Design (OD) is the open-source alternative. Same loop, same artifact-first mental model, none of the lock-in. We don’t ship an agent — the strongest coding agents already live on your laptop.

Type make me a magazine-style pitch deck for our seed round and OD’s interactive question form fires before the model improvises a single pixel. It stands on four open-source shoulders: huashu-design, guizang-ppt-skill, open-codesign, and multica.

31 Composable Skills

27 Prototype modes (web / mobile / dashboard / email / etc.):

• web-prototype, saas-landing, dashboard, pricing-page, docs-page, blog-post
• mobile-app, mobile-onboarding, gamified-app
• email-marketing, social-carousel, magazine-poster
• motion-frames, sprite-animation, dating-web, digital-eguide, wireframe-sketch
• critique, tweaks
• pm-spec, team-okrs, meeting-notes, kanban-board, eng-runbook
• finance-report, invoice, hr-onboarding

4 Deck modes: guizang-ppt, simple-deck, replit-deck, weekly-update

Skills follow Claude Code’s [[claude-ai/anthropic-skills-repo|SKILL.md convention]] — drop a folder into skills/ and OD picks it up. This is the same architectural pattern as the broader skills ecosystem (Superpowers, SuperClaude, oh-my-claudecode, agents).

72 Brand-Grade Design Systems

Each is a structured DESIGN.md payload extracted from a real shipped product. Categories:

• AI & LLM (10): claude, cohere, mistral-ai, minimax, together-ai, replicate, runwayml, elevenlabs, ollama, x-ai
• Developer Tools (~20): cursor, vercel, linear-app, framer, expo, clickhouse, mongodb, supabase, hashicorp, posthog, sentry, warp, webflow, sanity, mintlify, lovable, composio, opencode-ai, voltagent, [+others]
• Productivity (10): notion, figma, miro, airtable, superhuman, intercom, zapier, cal, clay, raycast
• Plus: Fintech (7), E-Commerce (6), Media (5), Automotive (6), Other (8), Starters (2)

This is the same architectural pattern as Refero Styles (real-shipped-product DESIGN.md extraction) and Neuform (free design system catalog), but locally hosted and modifiable rather than catalog-hosted.

15 Supported Coding Agents

Auto-detected by binary name on the user’s PATH:

CLI	Binary	Notes
Claude Code	claude	Primary integration target
Codex CLI	codex	OpenAI’s coding agent
Devin for Terminal	devin	Cognition’s autonomous engineer
Gemini CLI	gemini	Google’s coding agent
OpenCode	opencode	Open-source AI coding agent
Cursor Agent	cursor-agent	Cursor’s headless mode
Qwen Code	qwen	Alibaba’s coding model CLI
GitHub Copilot CLI	copilot	Microsoft’s CLI agent
Hermes	hermes	Nous Research’s self-improving agent
Kimi CLI	kimi	Moonshot AI
Kiro CLI	kiro-cli	—
Kilo	kilo	—
Mistral Vibe CLI	vibe-acp	Mistral’s auto-approve agent
DeepSeek TUI	deepseek	—
Pi	pi	—

Plus a multi-provider BYOK proxy that requires no CLI at all.

OD doesn’t ship an agent; it adapts to whichever the user already has — the architectural choice that makes it genuinely cross-vendor rather than Anthropic-only.

MCP Server (Read-Only)

OD ships a stdio MCP server that exposes the local project’s files to Claude Code, Codex, Cursor, VS Code, Antigravity, Zed, Windsurf, or any MCP-compatible client. Tools include search_files, get_file, and get_artifact.

The architecturally interesting bit (verbatim from README):

When the agent calls search_files, get_file, or get_artifact without a project argument, the MCP defaults to whatever project (and file) you have open in Open Design right now.

That means the agent reads exactly what the user is staring at — no export-reimport cycle. Setup is via Settings → MCP server panel. The daemon must run locally.

This is a different shape from Refero MCP (catalog-hosted, queries a remote design library) — Open Design’s MCP is your-local-project, exposing files for the agent to read directly.

BYOK Provider Surface

Multi-provider proxy that rejects loopback / link-local / RFC1918 destinations at the daemon edge:

• Anthropic (/api/proxy/anthropic/stream)
• OpenAI-compatible (/api/proxy/openai/stream)
• Azure OpenAI (/api/proxy/azure/stream)
• Google Gemini (/api/proxy/google/stream)

This decouples Open Design from any single vendor — skills can run on Claude Sonnet 4.6, Claude Opus 4.7, GPT-5, Gemini 2.5, or any compatible model.

Architecture

┌────────────────────── browser (Next.js 16) ──────────────────────┐
│  chat · file workspace · iframe preview · settings · imports     │
└──────────────┬───────────────────────────────────┬───────────────┘
               │ /api/* (rewritten in dev)          │
               ▼                                    ▼
   ┌──────────────────────────────────┐   /api/proxy/{provider}/stream (SSE)
   │  Local daemon (Express + SQLite) │   ─→ any OpenAI-compat endpoint
   │  /api/agents  /api/skills        │       (BYOK) w/ SSRF blocking
   │  /api/design-systems /api/projects/…
   └─────────┬────────────────────────┘
             │ spawn(cli, [...], { cwd: .od/projects/<id> })
             ▼
   ┌──────────────────────────────────────────────────────────────────┐
   │  claude · codex · devin · gemini · opencode · cursor-agent       │
   │  qwen · copilot · hermes · kimi · pi · kiro · kilo · vibe · deepseek  │
   └──────────────────────────────────────────────────────────────────┘

• Storage: SQLite at .od/app.sqlite for projects/conversations/templates/tabs; per-project files under .od/projects/<id>/.
• Preview: Sandboxed iframe via srcdoc + per-skill parser.
• Exports: HTML, PDF, PPTX (agent-driven), ZIP, Markdown.
• Media models: gpt-image-2 (Azure/OpenAI), Seedance 2.0 (ByteDance), HyperFrames (HTML→MP4).

Anti-AI-Slop Discipline

OD’s prompt stack composes a discovery-first system prompt:

• DISCOVERY directives (turn-1 form, turn-2 brand branch, TodoWrite, 5-dimensional critique)
• Identity charter (OFFICIAL_DESIGNER_PROMPT, anti-AI-slop, junior-pass)
• Active DESIGN.md (one of 72 systems)
• Active SKILL.md (one of 31 skills)
• Project metadata (kind, fidelity, speakerNotes, animations, inspiration IDs)
• Skill side files (auto-injected: assets/template.html + references/*.md)
• Deck framework directive (if applicable)

The discovery question form on turn 1 is the architectural defense against design-redirect cycles — the model asks before improvising. The 5-dimensional self-critique is a pre-emit quality gate. Both are readable TypeScript in apps/web/src/prompts/system.ts, which is itself a notable affordance: the prompt-stack composition logic isn’t a black box.

Position in the Ecosystem

OD sits alongside the other ai-web-design catalogs but occupies a structurally different slot:

	Catalog model	Runtime model	Open source	Multi-agent
MotionSites	Paid hero-prompt library	(catalog-only)	No	n/a
Refero Styles	Catalog + MCP	Cloud catalog	No	Any MCP client
Neuform	Free design-system gallery	(catalog-only)	No	n/a
Crafft	Node-based canvas + MCP	Cloud studio	No	Any MCP client
Open Design	Local-first runtime	Local daemon + spawned CLI	Apache-2.0	15 CLIs auto-detected

It’s the only entry that’s local-first, open-source, multi-agent, and Hermes-compatible. That makes it the natural target for users who want the Claude Design artifact-first experience without subscription lock-in or running everything in Anthropic’s cloud.

Key Takeaways

• Open-source alternative to Claude Design — same artifact-first loop, no closed-source / cloud-only lock-in. Apache-2.0, ~24.3k stars at ingest, TypeScript, Next.js 16 frontend + local Express + SQLite daemon.
• 15 coding-agent CLIs auto-detected from the user’s PATH — Claude Code, Codex, Devin, Gemini, OpenCode, Cursor Agent, Qwen, Copilot, Hermes, Kimi, Kiro, Kilo, Mistral Vibe, DeepSeek, Pi. OD doesn’t ship an agent; it adapts to whichever the user already has.
• 31 composable Skills following [[claude-ai/anthropic-skills-repo|Claude Code’s SKILL.md convention]] — 27 prototype modes (web, mobile, dashboard, email, magazine-poster, sprite-animation, …) + 4 deck modes (guizang-ppt, simple-deck, replit-deck, weekly-update). Drop a folder into skills/.
• 72 brand-grade Design Systems as DESIGN.md payloads — same architectural pattern as Refero Styles and Neuform, but locally modifiable. Categories span AI/LLM, dev tools, productivity, fintech, e-commerce, media, automotive.
• Read-only stdio MCP server exposes the local project to Claude Code / Cursor / Antigravity / Windsurf etc. — search_files, get_file, get_artifact. Defaults to the file the user has open. No export-reimport cycle.
• Multi-provider BYOK proxy — Anthropic, OpenAI-compatible, Azure OpenAI, Google Gemini. SSRF blocking at the daemon edge. Decouples skill execution from vendor model choice.
• Anti-slop machinery — turn-1 discovery form, identity charter (OFFICIAL_DESIGNER_PROMPT), 5-dimensional self-critique pre-emit gate, slop blacklist. The discovery question form is the architectural defense against design-redirect loops.
• Stands on 4 open-source shoulders — huashu-design, guizang-ppt-skill, open-codesign, multica. Composition over from-scratch authorship.
• Architecturally distinctive — the only entry in ai-web-design that is simultaneously local-first, open-source, multi-agent, and Hermes-compatible. Sits alongside AITmpl in the broader open-source-Claude-stack landscape.

Related

• Claude Design — the Anthropic closed-source surface OD positions against
• Claude Design Walkthrough (Paul Couvert) — working prompting method (transferable to OD)
• Claude Design — Six Agentic Patterns (Sam Witteveen) — architectural teardown of the original; same patterns OD’s prompt stack reproduces
• skills — canonical SKILL.md spec OD’s Skills follow
• Refero Styles — sister DESIGN.md catalog (catalog-hosted vs OD’s local)
• Neuform — free design-system gallery (alternative consumption model)
• Crafft — adjacent multi-modal creative studio with MCP
• MotionSites — paid hero-prompt library the catalog architecturally complements
• Hermes Agent — one of the 15 CLIs OD auto-detects
• Superpowers / SuperClaude / oh-my-claudecode — third-party Claude Code skill ecosystems with parallel ambitions
• AITmpl — adjacent open-source-stack catalog (BrainGrid / ClaudeKit)

Try It

1. Confirm prerequisites — Node ~24 and pnpm 10.33.x. Most macOS dev machines need a corepack enable && corepack prepare pnpm@10.33.x --activate step.
2. Clone and bootstrap:

   git clone https://github.com/nexu-io/open-design.git
   cd open-design
   corepack enable
   pnpm install
   pnpm tools-dev run web

3. Configure BYOK in Settings → providers. Anthropic key for Claude Sonnet 4.6 / Opus 4.7 is the natural starting point if you already pay for the API.
4. Confirm a coding agent is on your PATH — which claude should resolve. OD spawns it as spawn(cli, [...], { cwd: .od/projects/<id> }).
5. First prompt: something concrete like “make me a magazine-style pitch deck for our seed round.” OD’s discovery form will fire before any pixels generate — answer it before the model commits.
6. Wire the MCP server — Settings → MCP server panel. Then add it to Claude Code (claude mcp add open-design …). Now Claude Code sees the file you have open in OD.
7. Compare against Claude Design for a real project — do the same brief in both, evaluate output quality, exported artifact format compatibility, and total cost. The Claude Design Import feature lets you parse and continue editing Anthropic exports, so this is a low-friction A/B.

Open Questions

• Is there a multiplayer / Cowork-equivalent surface? Claude Design ships inside Cowork; the README describes OD as local-first / single-machine. Whether multi-user collaboration is on the roadmap or out of scope is unclear.
• Skill marketplace / registry. Is there a claudeskills.info-style registry for OD-specific skills, or do users clone repos? The README doesn’t surface a discovery surface beyond the bundled 31.
• Design-system extraction tooling. The 72 systems are presumably scraped or curated from real shipped products. Whether OD includes (or plans) an open-source extractor users can run on a target site, vs. systems being author-contributed only, is open.
• Hermes integration depth. The README lists hermes as a supported binary but doesn’t detail Hermes-skills auto-discovery. Whether OD’s skill loader picks up Hermes’s optional-skills/ tree automatically — a natural integration given both follow SKILL.md — is unconfirmed.
• Anthropic comparative cost. For a heavy user, what’s the OD-via-Anthropic-API cost vs. a Claude Design subscription? The break-even point would change the recommendation for cost-sensitive users — relevant to agency economics analyses.
• Self-critique fidelity. The 5-dimensional pre-emit gate is described in the README but the dimensions / rubric / scoring threshold aren’t enumerated. Reading apps/web/src/prompts/system.ts would resolve this.