Eight-plus years of client work gives you a very specific reflex: you see a build, you price it. Six years ago I was quoting $4,000 to $8,000 for a "custom" Shopify build that was mostly a stock theme plus three weeks of fighting the theme editor. Two years ago the same work had eroded to about $2,500. Then I spent an afternoon building a drop-ship product page — sticky purchase bar, conditional reviews, lifestyle carousel, a mobile layout I did not have to apologize for — inside one Antigravity workspace, for $59 a month of software. The build was materially better than those old projects. That afternoon did not just save money; it quietly ended a pricing model I built part of my career on.
This post is Shopify vibe coding from inside the work, not from the hype thread. The marketing layer around this stack is full of "anyone can launch a store in five minutes" claims that bury the three or four things that genuinely break if you do not know what you are doing. I will give you the stack, the honest math, and the failure points.
Why this works now when it didn't before
If you tried to connect an AI agent to Shopify in 2024 or early 2025, you probably had a bad time. The route was a community MCP server glued to the Admin API: products worked, orders broke, theme access depended on API-version luck, and breaking changes announced themselves as 500s on a client's live storefront.
That category got replaced on April 9, 2026, when Shopify shipped the official AI Toolkit — an open-source plugin that installs into Claude Code, Cursor, and Codex with two commands and carries skills covering the Admin API, Storefront API, theme files, analytics, and docs, maintained by Shopify itself. The difference is not connectivity; it is context. You are no longer asking a general model to remember Liquid syntax from training data. The toolkit knows which section to patch, which JSON schema the current API version expects, and which mutations require an explicit --allow-mutations flag before touching production. I did a full teardown of it in my Shopify AI Toolkit walkthrough; this post is what happens when you point that machinery at an actual store build.
The other half of the stack is Google Antigravity, the agent-first IDE Google launched in November 2025 — free in public preview, cross-platform, and crucially able to run Anthropic's Claude models rather than being a Gemini-only shop. The full bill of materials:
- Google Antigravity — the workspace: parallel agents, reviewable artifacts, browser recordings. Free.
- Claude Code — the coding agent, on the $20/month plan (real caveat below).
- Shopify AI Toolkit — the authenticated bridge into the store. Free.
- Shopify Basic — $39/month billed monthly, about $29/month annual.
- Node.js 20+ — prerequisite for the Shopify CLI the toolkit authenticates through.
- A fulfillment partner — for shipping, irrelevant to the build itself.
Total fixed software cost: $59 a month.
The $59 vs. $3,500 math, and where it bites back
The median SMB "custom" store build historically ran around $3,500: premium theme, freelancer or agency customization, first month of platform fees. The vibe-coding version is $59 plus six to twelve hours of your own time. On sticker price, that is an enormous compression. But sticker-price comparisons are the shallowest analysis available, and my quoting history obligates me to show you where the money you saved actually went.
You are now the developer — specifically, the QA layer. A real chunk of that $3,500 was the dev catching issues you did not know were issues: the 2MB hero image tanking mobile load times, the breakpoint that collapses on tablets, the checkout conditional that only fails for international addresses. Claude Code writes clean code; it does not volunteer that your LCP is dying unless you ask. The QA responsibility moved to you, undiluted.
Agent time is not free in practice. The $20 plan has rate limits, and a serious build session eats weekly quota faster than you expect. Hit the ceiling and your options are waiting for the reset or upgrading to a $100+ tier. Budget for the possibility.
The theme you skipped buying is the bar you have to clear. People buy premium themes for the forty pre-designed, polished sections, not because the free ones are bad. You can rebuild those sections by prompt — you are trading money for time, and you should be honest with yourself about which you have less of.
The CLI auth flow will cost you 45 minutes exactly once. URL handshake, one-time code, scope approval — a two-minute task on the happy path, cryptic errors off it. I pasted the code into the wrong terminal window twice. Now you will not.
The honest summary: you save $2,000-$7,000 and weeks of email round-trips in exchange for 8-20 hours of your time and the mental load of being your own QA. For most founders, obviously worth it. If your time genuinely bills higher on your core business, hiring the build out is still rational.
The build order that matters more than the tooling
After running this stack on real builds, my strongest opinion is about sequence, not prompts.
Start with the hero, and nothing else. The hero carries about 40% of the page's decision-making: purchase button, price, primary social proof, main image, one-line value proposition. But the deeper reason to start there is mechanical: the agent builds context as it works, and the context established by the hero — category, tone, design vocabulary, color system — propagates into every subsequent section. Start with the shipping-info block and your hero inherits bland context it then has to fight. My prompt pattern, near verbatim: reference screenshots of three heroes I like, explicit instructions on which element to take from which, "use my existing theme's color tokens, don't invent new colors," and the load-bearing final line — make the mobile layout primary; desktop adapts from mobile, not the other way around. AI-generated sections default to desktop-first and degrade downward; most store traffic arrives on phones; say it explicitly, every time.
Clone proven section patterns; do not innovate at the section level. Drop-ship conversion is a solved problem at the section level — press bars, comparison tables, review displays, what's-in-the-box modules. Screenshot the patterns from high-performing stores in your category and have Claude rebuild the pattern with your product, brand tone, and color system. Reference screenshots give the model a concrete target instead of a description to interpret, and it is unreasonably good at that specific task. Save your differentiation for product and brand, where it belongs.
Keep the image pipeline human. This is the drop-ship-specific insight I have not seen anywhere else, and it matters: your supplier ships a physical product. If an AI-generated hero image shows a subtly different version — wrong label orientation, wrong cap color, wrong proportions — you have set an expectation the box in the mail will not meet. That is a refund request per customer and an ad-account flag at scale. Generate code with the agent; source and export product imagery yourself (2048px long edge suits Shopify's transformers) and reference it by URL.
The last 15% happens in the theme editor. Section ordering, mobile preview, cart drawer, end-to-end checkout walk. An hour of manual QA that replaces the agency's QA layer. Skipping it is how afternoon builds become week-long refund investigations.
What Antigravity specifically adds
Everything above works in any Claude Code host. Antigravity earns its slot with two features. Its manager view runs multiple agents in parallel workspaces — hero, product module, reviews, policies each on its own agent, reviewed from one screen. The speedup is not faster agents; it is deleting the context-switch tax, the same economics that made me adopt parallel agents over git worktrees for application work. And its artifacts system means agents produce reviewable deliverables — plans, browser recordings of rendered sections — that you comment on like a doc. "CTA too small on mobile, minimum 48px tap target" typed on a recording becomes incorporated feedback without re-prompting. For once, the "vibe coding" label fits: it is closer to directing a small team than prompting a model. The broader philosophy of that shift is a thing I have written about in vibe coding apps without writing code.
One second-order effect worth naming: the stack compounds on a specific store. Early sessions need your palette, typography, and conventions restated constantly; a few sessions in, with the toolkit reading your actual theme files and the workspace memory accumulated, new sections land close to right on the first try far more often. That accumulated context lives in the workspace and CLAUDE.md, not the model — treat it as a project deliverable, because losing it resets the economics. It is also exactly why retainer-model agency work gets better margins from this stack every month while project-based pricing gets flattened by it.
What this stack still cannot do
Checkout is untouched, by design — Shopify owns that flow for PCI and fraud reasons, and the toolkit respects the boundary (Plus-tier checkout extensibility is a separate surface). Stock checkout converts fine; just know the limit. And SEO strategy remains human: the agent writes meta descriptions and schema on command but cannot tell you which keyword deserves targeting or whether your category's search volume justifies the effort. Judgment stays with you; the typing goes to the agent. That division, honestly, is the entire story of this stack.
For founders who would rather skip the QA learning curve entirely, I take this build on as client work — the same stack, with the mobile-first pass, the supplier-accurate image discipline, and the theme-editor sweep already accounted for. The formats are on my services page.
