We don't believe in inspiration galleries. We built one anyway — and it's a transfer protocol.

The honest confession.

We don't believe in inspiration galleries. We have spent fifteen years inside design teams, and not once did a Pinterest board ship a feature. Not once did a Mobbin scroll session translate into a working component. The reference industry, as it exists today, is a slot machine that pays out in dopamine and never in code.

And yet here we are, shipping one. The honest answer is that we built AI2 not because the world needed another gallery — it absolutely does not — but because the gallery is the wrong noun. What we built is a transfer protocol dressed in editorial clothing. The screenshots are bait. The structured tokens behind every page are the product.

A reference that doesn't compile is a reference that didn't help. The next decade of design tooling will be measured by what your agent can read, not what you can scroll.

What references actually do — and where the existing tools betray you.

Designers don't browse references the way the gallery industry pretends. We browse them looking for permission. Permission to use one accent color across an entire surface. Permission to ship a serif italic display at 56 pixels even though the brief said "clean." Permission to refuse a pattern everyone else has adopted, because Linear refuses it and Linear works.

Mobbin gives you a flipbook. Awwwards gives you a podium. Land-book gives you a directory. None of them give you the design system underneath. None of them give you the eight-color palette in OKLCH, the six-step type ramp, the two-easing motion vocabulary, the five accessibility pairs that pass APCA Lc above seventy-five. They give you a JPEG and a name and the polite suggestion that you figure the rest out yourself.

We are not interested in the polite suggestion. The reference is doing half the job; the half it isn't doing is the half that matters. The reference industry is a category that has decided, collectively, to stop short of the moment of transfer.

The agent moment is not coming. It already arrived.

Three things happened in the last eighteen months that the reference industry has not metabolized yet. Coding agents matured to the point where Cursor and Claude Code routinely write production components on first attempt. Design tokens crossed the W3C draft finish line and became a paste-ready interchange format. And the cost of running a serious extraction pipeline — full-page Selenium captures, font face inventory, motion library detection, dual-spec contrast audits — collapsed to the point that we can ship it as eighteen HTTP endpoints from a self-funded company.

Read those three sentences again. Each is, by itself, an interesting development. Together, they are a regime change. Together, they make every existing inspiration tool obsolete in the same way that Stripe made every existing payment processor obsolete: not by competing on features, but by changing what the workflow even looks like.

The new workflow is one call. URL goes in, design language transfer brief comes back. Your agent reads the brief, writes the component, ships the surface, ages the codebase forward. The screenshot is still there — but it's no longer the artifact. It's the proof of work. The artifact is the structured token tree, the curator brief, the agent prompt kit, the Tailwind config, and the promise that all four are extracted from the same source on the same Tuesday afternoon.

Eighteen APIs, one editorial promise.

When we tell people we shipped eighteen design APIs, the reaction is usually polite alarm. Eighteen sounds like feature creep. Eighteen sounds like a startup pitch deck pretending to be a roadmap. Eighteen sounds like, frankly, the kind of company we were trying not to become.

It isn't. Eighteen is what happens when you take a single editorial promise — "a curated reference is worthless if your agent can't act on it" — and translate it into every primitive that promise implies. Token Extraction, Screenshot, Brief Generation, Component Recipe, WCAG audit, Motion library detection: those are the six ways we honor the promise on the source side. Inspiration Catalogue, Detail, Related, Tag Cloud: those are the four ways we honor it on the read side. Agent Prompt Kit, Design Language Transfer, Multi-format Export: three ways the agent reaches for the work. Atomic Registry, Theme Generator, Starter Generator: three ways the work becomes a project. Webhooks, Storage, Analytics: three ways the operation runs at production scale.

Six plus four plus three plus three plus three. The arithmetic is the editorial line. Subtract any of them and you have stopped short of the moment of transfer. We will not stop short of the moment of transfer.

Why founding rates exist (and why we lock them for a year).

We are a self-funded company. There is no venture round to subsidize a free tier we cannot sustain. There is no acquihire we are positioning toward. There is no growth at all costs. There is, instead, a hundred-account founding cohort and the discipline of a company that has to be right about its pricing on the first try because the second try requires a real customer base.

The forty percent founding discount is not a marketing flourish. It is a contract. You believed in the work before the work had a public benchmark; we lock your rate for the year, and we lock the rate again on first renewal. After that, public pricing applies — but you have already paid us with the most expensive thing a small company can receive, which is conviction in advance of evidence. We honor that with a price you keep.

Founding rates are not a discount. They are a promise that the people who showed up early get to keep what they earned by showing up early.

What we refuse to build.

Every product is the sum of what it builds and the discipline of what it refuses to build. The refusals matter. They are how you read what a team actually believes, after the marketing copy is gone.

• We will not build a comments section. The gallery is editorial, not social. Comments would invite the metrics that destroy editorial.
• We will not personalize the catalogue. Fifteen curated items today, fifty curated items in a year, every item shown to every visitor in the same order. There is no recommendation algorithm to game.
• We will not run an ad network on top of agent exports. Sponsorship credit lines are editorial. Tracking pixels in your Tailwind config are not.
• We will not let the AI generate the curator brief. The agent assists, the editor signs. Every analysis on this site has a human signature even when the draft started in a model.
• We will not enter the screenshot-API price war. ScreenshotOne is a fine product at seventeen dollars a month. We are a different category at a different price for a different audience and we are at peace with that.

The invitation.

If you have read this far, you are the audience. Not the casual scroller — the one who has built systems, shipped components, audited for accessibility, written briefs nobody read, watched references rot in a Figma library because the team underneath them never stabilized. We made AI2 for you and the agent you are about to write code with.

Browse the gallery. Read a brief. Pull the Tailwind config. Hand the prompt kit to your model. Ship the component before lunch. Tell us what we got wrong. The feedback we received from the first hundred founding builders is already in the second iteration of the extractor; the feedback from the next hundred will define the third.

We are betting an entire company on the idea that the design reference, as a category, is about to be redefined by the agent. We are right. Come build with us while the rate is locked.