Open Design and the seat everybody waits on

The pitch deck a client actually sees, the landing page mockup that decides whether a campaign gets built, and the dashboard concept somebody needs before a build gets approved all come out of the same place in a small company. They come out of whoever holds the design seat. On Figma's pricing page that seat is $16 a month, $192 a year, for every person allowed to make something rather than look at it, and everyone else on the team gets view and comment access for free, which is a polite way of saying they get to wait. Open Design is an Apache-2.0 desktop app that moves the making part onto any laptop and hands the work to a coding agent you are probably already paying for. It shipped release 0.19.0 yesterday and the maintainers pushed code again this morning.

What the repo actually is

nexu-io/open-design describes itself, in the first line of its README, as the open-source Claude Design alternative. Claude Design is Anthropic's own conversational design tool, and on Anthropic's pricing page it is bundled into the $20 a month Pro plan rather than sold on its own. So the thing being copied here is not a drawing app. It is a product where you describe what you want and something else makes the file.

The interesting design decision is that Open Design does not ship a model. It drives a coding agent you already have. Claude Code, Codex, Cursor, Copilot, and by the README's own count more than twenty other command line agents, or a bring-your-own-key connection to anything speaking the OpenAI API shape, including a local Ollama instance. The app is a local Express daemon with a SQLite database that spawns whichever agent you point it at inside a managed project directory. Your existing subscription becomes the engine. That is the whole economic argument and I will come back to it, because it is also where the honest accounting lives.

What comes out the other end is files, not screenshots. Web, mobile, and desktop prototypes. Live dashboards. Presentation decks. Images. Video, through something the project calls HyperFrames. Exports run to single-file HTML with the assets inlined, PDF that knows it is a deck, PowerPoint, MP4, Markdown, and a zip of the lot. That matters more than it sounds. A generated PNG of a dashboard is a conversation. A PowerPoint file your operations lead can open and fix a number in twenty minutes before a board meeting is a deliverable.

The part I would look at hardest if I were evaluating this for a business is the design system handling. The repo bundles 151 design system packages and centers everything on a DESIGN.md file, which is the tool's memory of what your brand looks like. That is the difference between output that looks generated and output that looks like yours, and it is the thing most AI design tools quietly skip. Whether 151 of them are all good is a separate question and the answer is almost certainly no. One that matches your brand is enough.

Install is three doors. A signed desktop app for macOS on both Apple Silicon and Intel and for Windows x64, downloadable from the project site or straight off the GitHub releases page. A one-line od mcp install for people who want the agent integration and no interface at all. Or Docker Compose for a team that wants it on a shared box. Be clear on which of those you are: this is a download, not a service. There is no hosted URL where you sign up and start typing. For a non-technical person the desktop app is genuinely a double-click, and everything after the double-click is where the work is.

Traction is not in question. 85,228 stars and about 10,000 forks. Five releases in the four weeks between 0.15.0 on July 14 and 0.19.0 yesterday, the most recent adding image generation and the one before it a shared design workspace. Commits landed this morning. Apache-2.0, which permits commercial use with no copyleft trailing behind whatever you make with it, and that is a cleaner answer than most of what this series covers.

Why this lands differently for a small team

Run the seat math first, because it is smaller than people expect and that is the trap. Three people who need to make things, on Figma Professional, is $576 a year. Move to the Organization plan, where a Full seat is $55 a month, and the same three people cost $1,980. Neither of those numbers wakes an owner up at night. That is exactly why the seat count never gets revisited, and why the real cost hides somewhere else.

The real cost is the queue. In a twenty-person company there is one person with the seat and the skills, and that person has a job that is not making your mockup. So the request goes in a Slack thread. The campaign concept waits four days. The dashboard nobody can picture never gets approved, because approval requires a picture and the picture requires the queue. I have watched good internal ideas die at exactly that step, and not one of them showed up in a budget as a loss.

For the operations or revenue operations person, this is the same story as every other thing this site covers. You can already build a decent internal tool on a Friday afternoon. What you cannot do is produce the one image that gets the tool funded. That gap is silly and it is also real, and a local app that turns a paragraph into an editable dashboard mockup closes it without a ticket.

For the owner-operator the use is blunter. The recruiting flyer. The one-page quote sheet the field crew hands over. The deck for the bank. Those go to a freelancer or they go out looking like a Word document, and the freelancer is not expensive so much as slow, because you are one of eleven jobs that week.

Worth noting on the way past: Figma's own AI is metered now. A Professional Full seat carries 3,000 credits a month, Dev and Collab seats get 500, and the free Starter tier gets 150 a day capped at 500 a month. So the AI features you would compare this against are not sitting there unlimited on the plan you already pay for. They are a pool, and pools run out on the Thursday you needed them.

The honest take, which is where the free part gets complicated

Setup, in hours. The download is ten minutes. Then the app does nothing at all until a coding agent is on your PATH or an API key is in the settings. If you already run Claude Code or Codex, you are producing something rough inside an afternoon. If you do not, budget a full day for the first useful artifact and a second day to get a DESIGN.md that makes the output look like your company rather than like a template. Anyone telling you it is a ten-minute tool is counting only the download.

Hosting costs nothing, and that is a real difference from most of what gets covered here. No VPS, no Postgres to back up, no certificate to renew. It runs on the laptop. The bill did not disappear though. It moved to the model, and this is the part I would make somebody say out loud before they present this as a saving.

If you drive Open Design with your existing Claude subscription, you have not escaped a cap. You have pointed a new consumer at the one you already had. Anthropic's own pricing FAQ says activity across Claude on web, desktop, mobile, and Claude Code all draws from the same pool, on a rolling five-hour session window, with weekly limits stacked on top for paid plans. Design work is not cheap in tokens. Iterating a dashboard twelve times is a long session, and the next thing you hit the wall on is the customer email you were drafting. The alternative is bring-your-own-key, where the ceiling stops being time and becomes money: Sonnet 5 is $2 per million input tokens and $10 per million output. One deck is a few dollars at most. Twenty rounds on a dashboard with images is a number you should actually be watching, not assuming. There is also Open Design Cloud, the project's own paid top-up for people who want one balance across several models, which is a perfectly honest way for a free tool to stay funded and also the thing to notice before you write "free" on a slide.

What the paid products do that this does not. Figma is not really a drawing tool, it is a room. Several people in the same file, live cursors, comments pinned to the thing being argued about, version history, a handoff mode a developer trusts. Open Design is a local app on your machine. The README is straight about its own gaps: comment-mode surgical edits are only partially shipped, the panel for AI-suggested tweaks is not built, and the Figma migration plugin is at alpha. And the file format the rest of the industry trades in is still Figma's. If your agency sends you a design file, you are opening it in Figma, seat and all. Claude Design's advantage is a different one and it is not nothing: there is nothing to install, and the company that makes the model maintains the tool.

Who maintains this, and what happens if they stop. The org is nexu-io, 21 public repositories, with a product site behind it and several sibling projects that are themselves in the thousands of stars. The version number is 0.19.0, which is the maintainers telling you the truth. Five releases in four weeks is real velocity and it is also real churn, with 458 open issues and 345 open pull requests sitting there right now. Apache-2.0 means nobody can take the code back from you if the project stalls. What you would lose is the pace, and in a pre-1.0 tool the pace is most of what you are actually adopting.

Where the data ends up. Local-first is genuine here, and it is the strongest argument for this over anything hosted if your mockups contain real revenue numbers or a customer's name. Read the README's own wording though. Product analytics and session replay are consent-gated, and telemetry is described as scrubbed and always-on. Always-on is always-on, whatever it has been scrubbed of. Separately, bring-your-own-key means an API key is sitting in an app on the laptop of whoever installed it. That is a fine answer for one person. It is a bad answer for a team of eight with nobody rotating anything, and it is the same mistake this site has now watched businesses make with three different categories of tool.

What I would actually do with this

If one person on your team is the bottleneck for every visual that leaves the building, install this on the laptop of the person who keeps waiting on them, not on the designer's. The designer already has a seat and a workflow and will find this worse than what they have. The point is not to replace the person who is good at this. The point is to stop routing the rough draft through them, so that when something does reach the designer it is a revision instead of a blank page.

Keep the Figma seat for the person who genuinely lives in it. Drop the seats you bought so that two other people could occasionally make a slide, and see whether the local app covers those two. That is a $384 a year test on Professional and a $1,320 one on Organization, and it is reversible in an afternoon, which is the only kind of software test a small business should be running anyway.

Then watch the token spend for one month before you tell anybody you saved money. Not the seat line. The model line. Every tool in this category has moved the cost from a predictable subscription to a variable one, and a variable cost you are not measuring is not a saving, it is a surprise with a delay on it.

The seat was never the expensive part. The waiting was, and nobody has ever put that on an invoice.