E-commerce AI Stack

What's in this pack

This is the AI stack a working cross-border seller assembles after the first six months — the period where you've stopped pretending one ChatGPT subscription does it all and started asking which specific tool owns each stage of the funnel. Ten picks, organised into five layers so a small ops team can install them in the order the work actually happens.

Layer 1 — Product research and competitor intel. Apify MCP for scraping marketplace listings, review feeds, price history, and seller-rank changes. This is your top-of-funnel reality check before anyone writes a single product page.

Layer 2 — Listing copy and storefront work. The Shopify Expert agent for theme + Liquid + app development on the store side, plus the Content Marketer agent and the 34-skill Marketing Skills bundle for product-title, bullet-point, and description copy that survives the marketplace's character limits and reads like a human wrote it.

Layer 3 — Visuals. Together AI for fast hosted image generation (hero shots, lifestyle composites, ad variants), ComfyUI for the local node-based pipeline when you need controllable background removal, model-on-product compositing, or repeatable batch processing across hundreds of SKUs.

Layer 4 — Cross-border distribution. LibreTranslate as the self-hosted translation API for listings, support macros, and ad copy across the languages your storefront serves. Mautic as the email/lifecycle automation backbone — segmented by country, language, and lifecycle stage.

Layer 5 — Customer support and review intelligence. The Customer Support agent for triage + first-pass drafting, Chatwoot as the open-source unified inbox across email + live chat + Instagram + WhatsApp, and the Apify MCP feeding back competitor reviews and your own to close the product-research loop next quarter.

The pack deliberately overlaps in two places. Copy (Content Marketer agent + Marketing Skills bundle) — the agent is your generalist for first drafts; the skills bundle is the specialist library when generic copy stops converting. Images (Together AI + ComfyUI) — Together is the fast hosted API for one-off creatives; ComfyUI is the controllable local pipeline when you need 500 product photos with identical lighting and backgrounds.

Install in this order (research → listing → visuals → ads → support)

1. Apify MCP Server — 8,000+ Web Scrapers for Agents — start here. Before you write product copy, scrape the top 20 listings in your category: title patterns, bullet structure, review themes, price ladder, average rating drift. The cheapest competitive moat is knowing what is already on the page. Apify owns this layer because the actor library covers Amazon, Shopify, Etsy, AliExpress, eBay, and most marketplace surfaces without you maintaining scrapers.
2. Claude Code Agent: Shopify Expert — if you run on Shopify (or a Shopify-clone), this agent handles theme work, Liquid templating, app development, and Shopify Admin API integration. Storefront velocity is the first place a generalist LLM falls apart; a Shopify-specialist agent saves you the two-hour debug spirals on {% liquid %} blocks and metafield schemas.
3. Claude Code Agent: Content Marketer — generates the listing brief from a target keyword and a feature list. Outline, hook, audience, target keyword density, and platform-specific variants (Amazon bullets vs Shopify long-form vs Etsy story-led).
4. Marketing Skills — 34 CRO, SEO & Growth Skills for AI Agents — the specialist prompts you reach for when the Content Marketer agent's draft is generic. Headline-test skill, meta-description rewriter, landing-page CRO pass, ad-copy formulas. Use one or two per product, not all 34.
5. Together AI Image Generation Skill for Claude Code — the fast hosted image engine. Generate the hero shot, the lifestyle composite, and 4-6 ad-creative variants from one prompt + brand-system context. Predictable per-image pricing, no GPU to babysit.
6. ComfyUI — Node-Based AI Image Generation — install once you outgrow hosted-only. ComfyUI gives you a node graph for repeatable workflows: pure-white background removal, model-on-product compositing, SKU-batched generation with consistent lighting. Run it on your own GPU and the marginal cost per image is electricity.
7. LibreTranslate — Self-Hosted Translation API with No Rate Limits — translate listings, support macros, and ad copy across your storefront's languages without per-character vendor fees. Self-hosted means listings stay private until you publish; rate limits don't bite when you re-translate 5,000 SKUs after a brand-voice change.
8. Mautic — Open Source Marketing Automation Platform — email + lifecycle automation. Country-segmented welcome sequences, abandoned-cart flows in the customer's language, post-purchase review-request emails. Mautic owns this because you keep the contact data, you keep the cost predictable, and the open-source plugin ecosystem covers the cross-border quirks (multi-currency, multi-language, multi-storefront).
9. Claude Code Agent: Customer Support — first-pass support draft: classify the ticket, pull relevant order data, draft the response in the customer's language, escalate the ones the agent flagged uncertain. Pair with Chatwoot below for the human-in-the-loop layer.
10. Chatwoot — Open Source Customer Support & Live Chat — the unified inbox: email + live chat + Instagram DM + WhatsApp + Facebook + Twitter in one place. Self-hosted means customer chat transcripts don't live on a vendor's server, which matters when you're operating across GDPR + CCPA + China-cross-border regimes.

How they fit together (ASCII e-commerce pipeline)

  ┌── Apify MCP (competitor scrape) ──┐
  │  titles, reviews, prices, ranks    │
  └────────────────┬───────────────────┘
                   ▼
  ┌── Content Marketer + Marketing Skills ──┐
  │  (listing copy, bullets, descriptions)   │
  └────────────────┬─────────────────────────┘
                   ▼
          ┌── Shopify Expert ──┐
          │ (theme + Liquid +   │
          │  app + Admin API)   │
          └────────┬────────────┘
                   ▼
        ┌──────────┴──────────┐
        ▼                     ▼
  ┌── Together AI ──┐   ┌── ComfyUI ──┐
  │  (hero + ads)    │   │ (batch, bg,  │
  └─────────┬────────┘   │  composite)  │
            │            └──────┬───────┘
            └─────────┬─────────┘
                      ▼
          ┌── LibreTranslate ──┐
          │ (multi-lang listings│
          │  + ad copy + macros)│
          └─────────┬───────────┘
                    ▼
              ┌── Mautic ──┐
              │ (lifecycle, │
              │  abandoned, │
              │  post-purch)│
              └─────┬───────┘
                    ▼
        ┌── Customer Support Agent ──┐
        │ (triage + draft + escalate) │
        └─────────┬───────────────────┘
                  ▼
            ┌── Chatwoot ──┐
            │ (unified DM,  │
            │  human-loop)  │
            └──────┬────────┘
                   │
                   ▼
          (reviews / complaints) ──┐
                                   │
           feeds back to Apify ◄───┘
           for next quarter's brief

The two critical loops are: Apify → Content Marketer → Shopify Expert (research-grounded listing that ships in days, not weeks) and Chatwoot → reviews → Apify (customer + competitor signal feeds the next product brief). Get those loops right and "what should we list next" becomes a data question rather than a Friday Slack thread.

Tradeoffs you'll hit

• One-size-fits-all listings vs per-marketplace tuning. The Content Marketer agent can generate one master brief; the Marketing Skills bundle has the per-platform rewriter prompts. Amazon bullets are not Shopify product descriptions are not Etsy stories. Skipping the rewrite saves an hour and costs you 20% conversion — do the rewrite.
• Together AI vs ComfyUI. Together is hosted, predictable per-image cost, zero ops; ComfyUI is local, marginal cost ~free per image, but you own the GPU and the node-graph debugging. Rule of thumb: ≤50 product photos a week, stay on Together; ≥500 photos a week or strict brand-consistency requirements (pure-white background, identical lighting), invest in ComfyUI.
• Marketplace compliance vs creative freedom. Amazon listing rules cap title length, require specific category attributes, and ban promotional language in bullets. The same product copy that works on your Shopify storefront will get an Amazon listing suppressed. The Marketing Skills bundle has marketplace-specific rewrites; use them.
• LibreTranslate vs DeepL / Google Translate API. LibreTranslate is self-hosted, no per-character fee, no rate limit. DeepL is paid but the translation quality (especially for marketing copy in DE / JA / ZH) is materially better. Rule of thumb: technical bulletpoints and support macros go through LibreTranslate; hero headlines and brand-voice marketing copy go through a human translator with DeepL as the first pass.
• Seasonality. Q4 traffic is 3-5× normal for most consumer categories. Mautic abandoned-cart sequences that work in March will overwhelm a November customer base; throttle send volume, retest open rates monthly, and don't run a new ad-creative test during Black Friday. The Apify competitor scrape becomes a daily job in November, not weekly.

Common pitfalls (Amazon character limits, TM issues, image policy, multi-account risk)

• Amazon listing character limits and the silent suppression. Title 200 chars max (often soft-capped to 150 by category), bullet 500 chars, description 2000. If the Content Marketer agent's draft runs long, Amazon doesn't reject it — it indexes a truncated version and you only notice when conversion drops. Put a hard character-count check in your Shopify Expert agent's review pass before any listing publishes.
• Trademark and IP claims. Generating product copy that says "works like a [BrandName]" or images that include a Nike-style swoosh because the model has seen too many sneaker ads — both will get you a takedown notice and, repeated, an account suspension. Build a brand-block list (your competitors' names, common Western luxury brands, anything from the IP Accelerator brand registry) and run every draft and every generated image through it before publish.
• Image background policy violations. Amazon main-image rules require pure-white #FFFFFF background, no props, no text overlays, no badges. ComfyUI can absolutely generate compliant images — but only if your workflow enforces the background swap. Save a versioned ComfyUI workflow file for "Amazon main image" and a separate one for "lifestyle ad", don't reuse the same graph.
• Multi-account association risk. If you operate one Shopify store and one Amazon seller account and one Etsy shop from the same browser, same IP, and same Chatwoot inbox — Amazon's seller-policy team will eventually associate them. The penalty for associated accounts is suspension of all of them. Operate from separate browser profiles, separate IPs, and (most under-rated) separate Chatwoot inboxes per storefront.
• AI-translated listings that read like AI. LibreTranslate's output is grammatical but often flat. In ZH, JA, KO, and DE, flat copy reads as machine-translated and tanks trust. Always have a native speaker pass on the top 20 SKUs and on every headline that runs paid traffic.
• Chatbot pretending to be human. Many marketplaces (and most GDPR / CCPA regimes) require disclosure when a customer is talking to an AI. The Customer Support agent's first reply should say so; Chatwoot's automation rules should escalate to a human at the moment the customer asks. Hiding the AI loses you the trust and risks the policy violation.
• Review-mining scope creep. Apify can scrape competitor reviews to inform your next product brief. It can also scrape competitor customer emails, scrape behind paywalls, or scrape at a rate that gets your IP blocked. Stay on public marketplace surfaces, respect robots.txt, and use Apify's residential proxy rotation responsibly.