Open Design: The Claude Design Alternative I Actually Use

Open Design: The Claude Design Alternative I Actually Use

I almost didn't try Open Design.

It was a Saturday morning, my Codex CLI was open in one tab, Claude Code was running an agent in another, and the last thing I wanted was a fourth design tool to babysit. Anthropic had shipped Claude Design back on April 17, 2026 — I had played with it for about an hour the week it dropped, watched it generate a clean three-slide pitch deck, then closed the tab because the rate limits on my Pro plan kept stalling mid-prompt. Cool product. Walled garden. Move on.

Then a friend dropped a GitHub link in our Signal group and wrote four words: "This is the one." Repo name: nexu-io/open-design. Tagline: "Local-first, open-source alternative to Anthropic's Claude Design." 18.2K stars, five days old. I cloned it, ran pnpm install, opened localhost, and within twenty minutes I had a desktop newsletter landing page on screen — generated by Codex CLI, driven through Open Design's skill system, exported to a clean static HTML bundle I could hand to a developer.

That's when I closed the tab on Claude Design for good.

This isn't a takedown of Anthropic's product. Claude Design is a beautifully built tool inside the Claude apps, and for a designer who lives entirely inside Claude.ai it might be exactly enough. But it is paid, cloud-only, single-model, and rate-limited — and the moment you want to point your own coding agent at the canvas, plug in your own API key, or run it offline on a flight, the wall hits. Open Design tears that wall down. It's the same shape — left panel for prototypes, slide decks, media generation, design system picker, light/dark, the whole vibe — except it's free, runs locally on your laptop, and auto-detects whichever coding agent CLIs you already have installed.

I've been using it for ten days now. This is the honest version of what works, what doesn't, and who should switch.

What Claude Design Actually Is — and Why So Many People Hit Its Walls

Before I get into Open Design, here's the thing most posts skip: explaining what Claude Design is in a way that makes sense if you haven't tried it.

Claude Design is Anthropic Labs' first-party AI design product. It launched in research preview on April 17, 2026, powered by Claude Opus 4.7, and it lets you talk Claude into producing prototypes, slides, one-pagers, and short visual artifacts — all from inside the regular Claude interface. You describe a landing page. It generates a working prototype. You tell it to make a five-slide pitch deck for a SaaS launch. It produces something polished enough to put in front of investors. You ask it to apply your company's design system, and it can pull from your codebase or design files to keep the output on-brand.

The output is genuinely good. I've watched it produce hero sections that would have taken a freelance designer half a day. For non-designers — founders, PMs, marketers — it's a leap. Anthropic pitched it explicitly at "people who aren't starting from a design tool and need to get from an idea to something visual quickly," and that framing is right.

But here's where my Saturday-morning friend was right too. Claude Design has four hard limits that started costing me real time:

1. It's gated behind Pro, Max, Team, or Enterprise. That's $20/month minimum just to access the research preview, and $100-$200/month to actually use it without choking on rate limits. Fine if you're already paying. Awkward if you're trying to give a junior designer access without buying them a seat.
2. It runs Opus 4.7. Only Opus 4.7. No GPT, no Gemini, no Kimi, no local models. Whatever Claude is good at, you get. Whatever it isn't, you eat.
3. It's cloud-only. No offline mode. No local file access without going through the official MCP integrations. You can't point it at a folder on your Mac and say "look at these brand assets first."
4. Your work lives in Anthropic's infrastructure. Exports come out clean — PDF, PPTX, URL, or push to Canva — but the working document is theirs. If your account gets rate-limited mid-flow, you wait.

I'm not anti-Anthropic. I write about Claude Code constantly. But the super-agent race I covered last month made one thing obvious: in 2026, locking yourself to one model for any creative tool is a mistake. The frontier shifts every six weeks. Whichever provider was best in April is rarely best in June.

That's the gap Open Design walks straight into.

What Open Design Is — and Why "Local-First" Is the Whole Point

Open Design is an Apache-2.0 licensed project from Tom Huang and the team at nexu.io. The GitHub repo hit 18K stars in its first five days, which is the kind of trajectory you only see when a project nails timing. It launched while Claude Design was still in research preview — and while every developer I know was already screaming about the rate limits.

The pitch is clean: same surface area as Claude Design, but local-first, BYOK at every layer, and driven by whichever coding-agent CLI you already trust.

Here's what that actually means in the README, version-checked against the live repo:

• 19 composable skills. Not 31, despite what some early write-ups claim — the current count is 19. These are the design moves the system can chain together: generate, iterate, restyle, slice, export, and so on.
• 71 brand-grade design systems. Pick one before you prompt and the entire output inherits its tokens, type scale, and component conventions. Apple-clean, Stripe-clean, brutalist, editorial — the spread is wider than any first-party tool I've used.
• Up to 15 coding-agent CLIs auto-detected. On startup, the daemon scans your PATH for claude, codex, devin, cursor-agent, gemini, opencode, qwen, copilot, hermes, kimi, pi, kiro-cli, kilo, vibe-acp, and deepseek. Whichever ones it finds become candidate "design engines" you can swap from a model picker.
• BYOK API mode if you have no CLIs installed. Plug in an OpenAI, Anthropic, or compatible key directly.
• Sandboxed preview, HTML/PDF/PPTX/MP4 export, SQLite persistence, and an agent-to-agent file-access daemon — the daemon reads and writes files, runs commands, and serves design files to whichever agent you're using.
• Media providers for speech, TTS, and image generation, configurable per project.
• Web prototypes, desktop, mobile, slides, images, videos, and HyperFrames — the full output surface mirrors Claude Design and goes further on video and HyperFrames.

The thing that hits hardest the first time you open it: the UI mirrors Claude Design almost beat-for-beat. Left panel with prototyping, slide decks, media generation, templates. A reasoning-effort slider. Light/dark toggle. Design system picker right at the top. There are even visual "pet" companions for workspace navigation — a small detail that tells you the team studied the Claude Design experience hard and then asked, "What if we kept the warmth and ditched the wall?"

That's the whole bet. Same vibe. Different ownership model.

I want to show you what that looks like in practice — but first, the install, because this is where most "open-source alternative" projects die.

Installing Open Design Without Setting Your Laptop on Fire

This is the part where I almost gave up on Saturday.

Not because Open Design is hard to install — it's actually one of the cleanest pnpm setups I've shipped this year — but because the README assumes you read carefully. I didn't. I tried pnpm dev first (a removed legacy alias), got a confusing error, and almost closed the repo. Don't do that. Read the QUICKSTART.md. It's short. It works.

Here's the sequence that got me running on macOS Sonoma in roughly 18 minutes, including the agent CLI configuration step.

Step 1 — Get Node 24 on your machine.

Open Design needs Node 24+. If you're on nvm, that's nvm install 24 && nvm use 24. If you're on fnm, the equivalent. If you're still on Node 18 because some other project pinned you there, just shell-scope it for this project — you don't need to nuke your global Node.

Step 2 — Enable Corepack so pnpm matches the repo's pin.

corepack enable

The repo pins [email protected] through the packageManager field. Corepack reads that and uses the right version automatically. This is what saves you from the "works on my machine" install drift that haunted JS tooling for a decade.

Step 3 — Clone and install.

git clone https://github.com/nexu-io/open-design.git
cd open-design
pnpm install

The install pulled down a chunk of dependencies — the project bundles a daemon, web layer, desktop shell, and the skills runtime — but it finished without a single yellow warning on my machine, which is more than I can say for half the AI repos I touch.

Step 4 — Run the dev server.

This is the command that actually launches the thing:

pnpm tools-dev run web

That starts the daemon plus the web layer in the foreground. You'll see the daemon boot, scan your PATH for coding-agent CLIs, and report which ones it found. On my machine it picked up claude, codex, gemini, cursor-agent, and opencode automatically — which was exactly what I wanted.

If you want the full stack (daemon + web + desktop) you run pnpm tools-dev with no args, but the web target is the right one for the first walkthrough. macOS, Linux, and WSL2 are the primary supported paths. Windows native should work for most flows but the maintainers recommend WSL2 as the safer baseline.

Step 5 — Open localhost and enable the core pack.

The terminal will print a localhost URL — for me it was http://localhost:5173. Open it. The first thing the app asks is to enable the core pack (the bundle of skills + design systems). Confirm. On Win/macOS this takes about thirty seconds.

Step 6 — Configure your agents and providers in the daemon UI.

Inside the app, there's a Settings → Runtimes panel. This is where the magic lives. You'll see every CLI agent the daemon detected, ranked by which one is healthiest. You can:

• Pick which CLI runs as your default "design engine."
• Drop in a BYOK API key for any provider not represented as a CLI.
• Wire up media providers — TTS, speech, image generation — separately. I plugged in an ElevenLabs key for narration and an OpenAI key for image generation. Both worked first try.

If none of your local CLIs are detected (say you're on a fresh Mac), BYOK mode picks up the slack. You're never stuck.

The whole flow is the kind of thing I would've expected to take a weekend. It took less time than installing the latest Adobe Creative Cloud update.

The Five-Minute Test That Made Me Switch

Here's the project that convinced me.

I needed a landing page mockup for a newsletter I'm launching in June — high-fidelity, desktop-first, sleek, and ready to hand off to a developer. Normally I'd open Figma, sketch wireframes, then either commission or build it myself in Tailwind. Instead, I opened Open Design and ran it through the standard new-project flow.

Project type: New design. Form factor: Desktop / web. Fidelity: High (not wireframe). Engine: Codex CLI. Design system: Editorial — the one closest to my actual brand. Reasoning effort: Medium.

Then I gave it a real prompt — not a clean demo prompt, the kind I'd actually type:

"Newsletter landing page for an AI engineering newsletter called Signal. Audience is senior developers and tech leads. Hero with a strong opinionated tagline, three short value props, recent issue previews as cards, a clean email capture, and a footer with social. Sleek, editorial, no emojis, no gradients. Light mode only."

Codex picked it up, asked two clarifying questions through the skill system — "What's the article angle for the latest issue card?" and "Do you want the email capture above or below the issue cards?" — and went to work. The reasoning slider was visible in real time. About 90 seconds later, the right panel rendered a working static prototype. Real typography. Real spacing. The hero tagline hit. The card grid was on the 12-column rhythm I would've picked myself.

I commented inline on three things: tighten the tagline by four words, swap the card hover state to a subtle border instead of a shadow, and increase the footer padding. Codex regenerated the affected sections — not the whole page. That alone is worth the install. Claude Design regenerates more aggressively, which is fine until you're three iterations deep and don't want to lose your manual tweaks.

The final export gave me an HTML bundle plus a PDF preview plus a screenshot I could drop in a Notion brief. Total time from pnpm tools-dev run web to "I can hand this off": about six minutes.

That's the moment my Claude Design tab closed for good.

Where Open Design Goes Beyond Claude Design

After ten days, four of the differentiators stopped feeling like marketing and started feeling load-bearing.

Multi-model is not a feature. It's freedom.

I've stopped thinking of "which model" as a setting and started thinking of it as a choice per task. Codex gave me the cleanest landing-page output. Claude Code, when I switched the engine, wrote better long-form copy inside the slide decks. Gemini was the fastest at ideation passes when I just wanted six layout variations to choose from. Open Design lets me match the model to the job. Claude Design can't do that, by definition.

If you've already read my piece on the Codex-Claude Code two-agent workflow, the mental model is the same: different agents, different strengths, route the work. Open Design just brings that exact pattern into the design surface.

The agent-to-agent file-access daemon is the quiet hero.

The daemon runs underneath the UI. It reads and writes files, runs commands, and — the part that matters — serves design files to other agents on your machine. So when I finished the newsletter mockup, I switched to Claude Code, asked it to generate the React + Tailwind version of the page, and pointed it at the design files Open Design had just written. Claude Code read them like any other source file. No copy-paste. No screenshot prompting. The handoff was literally one path.

This is the kind of architecture that took me three months of building my own agent teams to appreciate. Tools that don't talk to each other are the bottleneck of 2026 AI workflows. Open Design just talks to whatever else is on your laptop.

71 design systems is more than I expected to ever use — and yet I keep needing them.

I assumed the design system picker was a checkbox feature. Pick "Apple-clean," move on. Wrong. The systems span editorial, brutalist, glassy, neumorphic, claymorphic, retro-print, technical-doc, e-commerce, and more. Switching the system before generation shifts the entire output — typography, spacing, component shapes, color reasoning. I now think of it the way I think of a design system workflow in Figma: it's the foundation, not a skin.

Export is real, not gated.

HTML, PDF, PPTX, ZIP, MP4. The exports are local files on your machine. No cloud round-trip. If you've ever lost a Claude artifact because a session timed out, you know why this matters.

Where It Falls Short — Because No Tool Is Magic

Open Design is good. It's not perfect. Three honest gaps after ten days:

The drawing tools aren't there yet. The roadmap says they're coming. Right now, if you want to draw — sketch a freeform shape, pull a curve, manually move a node — you're stuck commenting and re-prompting. Claude Design has the same limit. Figma still wins on direct manipulation.

Setup is friendlier than most open-source projects, but it's still developer-only. If you're a non-technical founder who wants to "just open the app," this is not for you yet. The desktop build will fix that. Until then, you need a terminal and basic comfort with pnpm.

The skill system is powerful but undocumented for new users. The 19 skills compose into long chains, and the system is smart enough to pick reasonable defaults. But the moment you want to write a custom skill, the docs are still thin. The team is shipping fast, so this might be solved by the time you read this.

Some skill chains take real time at high reasoning. A complex desktop prototype with high reasoning effort is not a 90-second job — it can run two to three minutes. That's still faster than most human design loops, but if you're used to instant chat output, recalibrate expectations.

I'm flagging these because I don't trust glowing reviews of any tool, and I'd want someone to flag them to me. None of them are dealbreakers. All of them are fixable with engineering time the project clearly has.

Who Should Switch — and Who Shouldn't

After ten days I have a clean mental model for who Open Design is for.

Switch if:

• You're a developer or technical operator already running CLIs like Codex, Claude Code, or Gemini.
• You care about owning your work locally — your files, your keys, your machine.
• You want to use the right model for the right task instead of being locked to one provider.
• You're building client work and need exports that don't depend on a third-party account staying alive.
• You've been hitting Claude Design's rate limits.

Don't switch (yet) if:

• You live entirely inside the Claude.ai web app and have no other AI tools installed.
• You're a non-technical user who needs a polished installer and zero terminal contact.
• You need direct manipulation drawing tools today — not on the roadmap, today.
• Your team's shared Canva integration is mission-critical (Claude Design exports to Canva natively; Open Design exports to local files you'd then import).

For me, the math was simple. I was already running four agent CLIs locally. I was already paying for the API keys behind them. I was already exporting design files into Claude Code anyway. Open Design just turned that messy stack into a clean surface.

If you'd rather have someone wire this kind of multi-agent design and dev workflow into your team without the install pain, I take on those engagements through my Fiverr profile — usually a one- or two-week setup that ends with your team running their own Open Design + agent stack on their own machines.

The Bigger Pattern Open Design Is Part Of

Open Design is not an isolated event. It's the latest move in a pattern I've been tracking all year: the open-source community catching up to — and then leapfrogging — the closed AI products inside six to eight weeks of launch.

Claude Code shipped. OpenCode and the wider open agent CLI scene followed. Anthropic shipped Skills. The community shipped open Skill marketplaces. Anthropic shipped Claude Design on April 17. Open Design crossed 18K stars by April 22. The compression is the story.

What's different about Open Design is that it didn't just clone Claude Design's surface. It rethought the runtime. Claude Design runs Opus 4.7 on Anthropic's servers. Open Design runs whatever you have, wherever you have it. That isn't a feature delta. That's a different theory of how creative AI tools should work in 2026.

If you've spent 2025 watching the closed players ship beautiful demos and 2026 watching the open community ship beautiful tools, you already know which way the wind is blowing. Local-first + BYOK + multi-model isn't a hobbyist niche anymore. It's becoming the default expectation for any AI tool serious operators rely on. The only question is which closed product gets cloned this month.

This week, it was Claude Design. Open Design beat the clone curve.

Let's Work Together

Looking to build AI systems, automate workflows, or scale your tech infrastructure? I'd love to help.

• Fiverr (custom builds & integrations): fiverr.com/s/EgxYmWD
• Portfolio: mejba.me
• Ramlit Limited (enterprise solutions): ramlit.com
• ColorPark (design & branding): colorpark.io
• xCyberSecurity (security services): xcybersecurity.io