Claude has been the AI tool I reach for most days and that's probably not changing any time soon. Beyond the chatbot itself, Claude Design has also become a welcome addition to my workflow ever since Anthropic launched it a few months ago. I didn't even need to pay extra for it because it came along with the Pro plan I already had for general Claude use. So I was a little lucky on that front.

But I'm also one of those people who can't help testing alternatives to the tools I'm already happy with, especially when AI tooling is moving as quickly as it is right now. Pro is a shared pool too and design work eats through it way faster than most other things I do in Claude, so when an alternative came across my desk that lets me design either pay-as-you-go via API or completely free with a local model, it was an easy thing to look into. Open Design is what I ended up testing. And it ended up being more than I expected going in…

The open-source version of Claude Design

Same artifacts idea, very different shape

Open Design dropped almost immediately after Anthropic announced Claude Design and was created as a direct open-source alternative. It runs as a native desktop app on macOS, Windows, and Linux, which is already pretty different than Claude Design, which is cloud-only and tied to a Pro/Max/Team plan.

The artifact approach is the same though. You describe what you want and the app streams a real HTML/CSS/JS deliverable into a sandboxed preview pane, with HTML, PDF, PPTX, and even MP4 as the export options. Open Design doesn't actually ship its own model, which I think is the whole point. It shells out to whatever coding agent CLI you have on your machine, like Claude Code, Codex, Cursor, OpenCode, plus a bunch of others. So the model layer is completely decoupled from the app and that's where the cost flexibility comes from.

It ships with 250+ skills (landing pages, dashboards, decks, mobile apps, basically every common design surface) and 140+ design systems baked in, and all of them are plain Markdown files anyone can edit. Version 0.9.0 also introduced their own model router called AMR for people who don't want to set up a CLI. You top up a wallet, pick a model, and that's basically it. And the option that really pulled me in was BYOK with any OpenAI-compatible endpoint, which means you can also point it at a local model running through something like LM Studio.

Going the local route first

It wasn't smooth sailing

Honestly my first reaction when I opened the app was a little bit of panic. The runtime picker asks you to choose between AMR, a local coding agent CLI like Claude Code or Codex or Cursor, or BYOK. I don't use any of those CLIs and the idea of installing one and configuring it through a terminal made me want to close the app altogether, so local LLM through BYOK was the natural pick for me.

You just pick OpenAI from the provider dropdown and point it at the LM Studio localhost endpoint (http://localhost:1234/v1) since LM Studio exposes an OpenAI-compatible API anyway. Type any string into the API key field because it doesn't actually validate, hit Fetch Models, and your loaded LM Studio models populate the model dropdown.

At first I went with Gemma 4 E4B because it's been my go-to lately and it's also optimized for analyzing UI design with vision, so I thought it would be a perfect fit, but I couldn't be more wrong. The result was nothing usable and was just the raw skill template echoed back as broken XML in the chat, which made sense after I looked into it because despite having strong vision, Gemma isn't really tuned for the kind of multi-step agentic work this app expects from the model.

So I swapped to Qwen 3.5 9B and finally got somewhere. It ran through Open Design's discovery form properly, which is this side panel that asks for a few more details after your initial prompt like tone, audience, and project context. Then it streamed actual HTML and CSS into the preview. The result wasn't going to win any design awards but it hit every section I'd asked for and even nailed the tech/utility monospace vibe I'd picked in the discovery options.

The catch is that each generation took several minutes. Qwen runs at about 13 tokens per second on my 8GB GPU which is fine for chatting but slow for streaming out a whole HTML artifact, and the plugins also didn't seem to actually take effect when I tried to apply them to local generations. So a fair result for free but slow and limited, and at that point I started wondering what it'd look like with a "proper" model behind it.

Switching to Anthropic's API

It came together quite nicely

I happened to already have an Anthropic API key with some credit on it from a previous tool I'd been testing, which made the switch take about two minutes. Once the key was in, Open Design's BYOK tab pulled my available models automatically and I went with Opus 4.7 since this was specifically for testing visual output.

The difference was immediate. What took Qwen something like five minutes was done in well under one. And the output was just better in every direction, with tighter typography, smarter section hierarchy, and a clean editorial-tech mix I'd actually had in mind. For anyone considering going local but whose hardware can't really support a code-tuned model, paying for the API works out cheaper than a Pro sub if your usage is intermittent, and the speed difference alone makes it worth it.

The plugins also finally worked this time. Plugins in Open Design are basically community-contributed style biases that you attach to your prompt as a chip, and they shape the aesthetic of the generation without you having to spell out the visual direction yourself. There's also a sketch mode where you can rough out a layout on a canvas inside the app and the agent will turn that drawing into a real mockup, similar to how Claude Design does it. It's basically a place for lo-fi wireframing before moving onto creating the prototype.

The one part that didn't get better with the API swap was the manual editing. Open Design lets you click directly on elements in the preview to make targeted changes which sounds great in theory, but the edit box jumps around with the cursor that makes it hard to actually land on what you want to edit. I assumed it was something my local model was doing but it persisted with Opus too. Turn out, they seem to be upfront about this on their roadmap. Comment-mode surgical edits are listed as partially shipped and the tweaks panel UX is listed as not yet implemented.

The actual portfolio it spat out was honestly something I could see myself using for real, and I opened it in my browser to just keep poking at it. The HTML and CSS are real and properly responsive, and with a bit of refinement I think it could be something I'd actually publish. I just don't have a domain to put it on so for now it lives as a local file.

More viable than I expected

I'm not ditching Claude Design, it comes bundled in my subscription anyway so it's not like I can get rid of it. But Open Design has a real case for itself. Free if your local model can handle it, cheaper than Pro if you go API pay-as-you-go for intermittent design work, and it has things Claude Design doesn't really handle as well, with the more diverse and integrated plugin system at the top of the list. It still has some quirks, but it's worth a look either way.