Design tooling has been getting more attention on the AI roadmap lately, and I think a lot of it comes down to vibe-coding making its way into design. Everyone doing Claude Code is one click away from Claude Design, and I reckon that’s a big part of why more people are wandering into design at all. Which is a good thing. Before Claude Design existed, I was already messing around with Figma Make, and there have been no shortage of options after that.
The indie, local, and open-source side of AI and design is where it’s worth poking around for hidden gems, though. OpenPencil and Open Design were my entry into hooking local LLMs into design work, and both are still on my PC. But I recently came across another one that does something similar. It’s called Paper, and although it’s a hosted tool and not open source, it lets you plug in a local model via MCP and design entirely on your own hardware.
Want to stay in the loop with the latest in AI? The XDA AI Insider newsletter drops weekly with deep dives, tool recommendations, and hands-on coverage you won’t find anywhere else on the site. Subscribe by modifying your newsletter preferences!
What Paper actually is
A code-native design tool that came out of nowhere
Paper was founded in 2024 by Stephen Haney, who was previously a co-creator of the Radix UI component library. It’s a hosted SaaS tool, not open source and not self-hostable, currently in open alpha with a desktop app and a web version. But don’t let that put you off before giving it a shot, because the premise is what makes it interesting… Every element you place on Paper’s canvas is real HTML and CSS. So there’s no proprietary file format and no export layers; the design is in code. This means that AI agents can read and write to the canvas without a translation step in between.
The free tier is generous for exploring – you get the full canvas and 100 MCP calls per week, which is what I’ve been using thus far. Pro is $20 a month (or $16 for annual) and bumps that up to a million MCP calls per week. There’s also an Organizations tier coming for teams. If you’re a solo hobbyist or a designer-developer poking at agent workflows, the free plan is honestly fine. If you’ve read any of my design or AI tool coverage, you’ll know I’m a bit stingy with subscriptions and will always recommend the free tier indefinitely.
Feature-wise, it’s mostly familiar territory if you’ve ever used a UX/UI tool like Figma or Penpot. There are frames, vector tools, text, real flex layout with on-canvas gap and padding handles, components, tokens, and an OKLCH color picker. It’s nothing crazy or special, really. The only two things that stand out are the shader library (GPU-accelerated stuff like mesh gradients, liquid metal, and fluted glass that run natively on the canvas), and the token system, which I’ll come back to later because that’s where the local LLM story actually gets good.
Hooking up Paper to a local LLM
There are a few routes: mine was fastest
Paper’s whole thesis about AI is that agents should be first-class citizens on the canvas, not bolted on afterwards. The MCP server ships with 24 tools that are bidirectional, meaning an agent can read your file and write to it. Which is actually better than Figma’s read-only setup, and I think it’s why Paper feels so agent-native compared to everything else. Their pitch to the designer-developer crowd is basically that the canvas already speaks the language of the browser, so an LLM reading it isn’t translating between formats, it’s reading real HTML. That also means Paper uses fewer tokens per task than Figma MCP does.
My route was LM Studio. Paper Desktop starts an MCP server at http://127.0.0.1:29979/mcp whenever a file is open, so I just added that URL to LM Studio’s mcp.json, toggled it on in the chat panel, loaded my model (Qwen 3.5 9B), and that was it. Took maybe five minutes.
But LM Studio is only one option. Paper’s docs cover Cursor, Claude Code, Codex, VS Code with Copilot, OpenCode, and Antigravity as officially supported hosts. And if you’re on Ollama specifically, there are MCPHost and ollmcp as bridge tools that plug your local model into any MCP server. So you’ve got options.
Testing Paper on my own hardware
Impressive, a little ugly, and eventually useful
Qwen 3.5 9B was my model for this because it’s been the best performer for local design I’ve tested in other tools so far (and that I can run on my hardware). It also has native tool calling built in, and Alibaba specifically calls it out as an agentic model, so it handles MCP’s structured JSON output more reliably than Gemma does.
My first real test was a reusable icon button set, which in UX is basically a component library. Qwen delivered a proper 4×2 grid, each button with its own editable node with a child element inside, exactly the structure you’d want if you were going to reuse them. So structurally, it did the job. Aesthetically, though, the results were just ugly, for lack of a better term. I gave it emoji glyphs to save tokens and Windows renders those as 3D blobs. But that was kind of on me for being a little vague in the prompts.
The second attempt was much better. I specified a thin-stroke sage green line art aesthetic to match a meditation app I’ve been using as practice in Figma, and it delivered. Which tells you something about how much of this is the prompt doing the work and how much is the model. Qwen isn’t guessing your design taste, it’s executing what you tell it. So I definitely recommend refining your prompts.
Local has real ceilings though. Qwen advertises a 262k context window, but I can practically push it to around 32k on my hardware because VRAM caps how much context you can actually load. And every MCP tool call carries the full Paper tool schema back into the window, so tool-heavy sessions burn through it fast. I hit the ceiling more than once and had to restart chats.
The tokens use case is where this all pays off though. In UX design, tokens are named reusable values for colors, spacing, radius, typography, and so on. So instead of hardcoding a hex value everywhere, you reference sage-primary and every use updates when you change the token. It’s a simple version of a design system. And in Paper specifically, tokens can only be created via the MCP right now, there’s no UI for it. Which means the AI hookup isn’t just a nice-to-have, it’s the actual mechanism…
I had Qwen generate 29 tokens for me based on my meditation app aesthetic – 8 colors, 6 spacing steps, 4 radii, and a typography scale – and they’re all real, viewable via File > Theme > Copy Theme, and applicable to any element through the token dropdown.
One of my favorite AI-first tools right now
Paper wasn’t on my radar until recently, even though it’s been around for a few years now. It’s one of the first design tools I’ve used where AI feels properly integrated into it rather than tacked on – and running it with a local model just gives you more ownership of the design process. My hardware is still a bottleneck, but for what I’ve been doing with it, Qwen 3.5 9B plus Paper has been surprisingly workable.
