Catalog Attribute Normalizer

Taxonomy-grounded catalog attribute normalizer — verified against the real Google Product Taxonomy, so it catches the plausible-but-wrong category IDs a generic LLM invents. Messy multi-source catalogs — titles, descriptions, images — in; consistent attributes, units, and grounded category mappings out. For merchant-ops teams and feed-tool developers wrangling a Shopify export, a supplier CSV, and a marketplace scrape that each spell "size" or "material" differently.

What it does

• Accepts a batch of raw products (title, description, raw_attributes), returns normalized attributes (canonicalized sizes/colors/units) plus category mappings for whichever taxonomies you request (Google Product Taxonomy, Shopify, Amazon).
• Works as an MCP server (normalize_catalog tool) or a plain HTTPS API.

Example (MCP tool call)

{
  "tool": "normalize_catalog",
  "input": {
    "products": [{ "title": "...", "description": "...", "raw_attributes": { "size": "Lrg", "Color": "Navy Blue" } }],
    "target_taxonomies": ["google", "shopify"]
  }
}

{
  "results": [
    {
      "schema_version": "2.0",
      "source_title": "...",
      "category_paths": {
        "google": { "path": ["Apparel & Accessories", "Clothing", "Shirts & Tops"], "leaf_id": null, "confidence": 0.9 },
        "shopify": { "path": ["Apparel & Accessories", "Clothing", "Tops"], "leaf_id": "aa-3-1", "confidence": 0.85 }
      },
      "attributes": {
        "size": { "value": "L", "provenance": "canonicalized" },
        "color": { "value": "navy", "provenance": "canonicalized" }
      }
    }
  ]
}

attributes is keyed by a controlled vocabulary (size, color, material, gender, sleeve_length — unrecognized keys are dropped, not passed through under a model-chosen name). Each value's provenance is "canonicalized" when it came from your own raw_attributes input for that product (deterministic cleanup only, no recall) or "extracted" when the model inferred it from the title/description and it wasn't in your input — treat "extracted" values as a suggestion, the same way you'd treat a low-confidence category_paths entry.

Attributes are reliable by construction — canonicalizing values already in your input, not recall. Category classification is retrieval-grounded, not recalled from memory: candidates are retrieved from the real, current Google and Shopify taxonomy files and offered to the model as suggestions, so a leaf_id almost always names a node that actually exists — measured against a 12-product evaluation set, each product checked against both the Google and Shopify taxonomies: 22 of 24 checks (91.7%) exact path + leaf-ID matches. Still treat confidence and a null leaf_id as "worth a quick check," not a guarantee — a leaf_id is only ever returned when the path independently verifies against the real taxonomy file, so a null there is an honest "check this" signal, never a fabricated ID. Amazon has no comprehensive public taxonomy reference file to retrieve candidates from, so it stays best-effort (recall from memory) rather than retrieval-grounded.

Pricing

The free tier is self-serve today; paid plans are opening soon.

• Free — 500 products/month, no card required. Self-serve today.
• Pay-as-you-go (opening soon) — $0.01/product, no minimum.
• Pro (opening soon) — $29/month for 5,000 products (~$0.0058/product effective).

Want a paid plan now? Register your interest and we'll set you up first.

What counts as one unit: one product, not one call. A normalize_catalog call carrying 40 products uses 40 of your monthly allowance; batching is a convenience, not a discount. A batch is all-or-nothing — if it is larger than your remaining balance, the whole call is refused with 402 (the body names required and remaining) and nothing is classified, so you are never billed for a partial result that looks complete. A call that fails outright costs nothing.

Getting a key

The MCP server is live at https://catalog-normalizer.acjlabs.com/mcp (the alternate https://acjlabs-catalog-normalizer.acjlabs.workers.dev/mcp address reaches the same deployment and keeps working, so existing configurations need no change). Free-tier keys self-serve — POST /v1/signup with { "email": "you@example.com" } returns your key directly in the response, good for 500 products/month, no card required. Store it immediately — it is shown once and cannot be recovered if lost (re-signup for a new one). See https://acjlabs-catalog.pages.dev for the same steps plus pricing and a comparison against alternatives. Paid keys provision the same way via a one-time claim link once Pro/pay-as-you-go billing is live.

Connecting it to an MCP client

With Claude Code:

claude mcp add --transport http catalog-normalizer \
  https://catalog-normalizer.acjlabs.com/mcp \
  --header "Authorization: Bearer YOUR_API_KEY"

Any other MCP client that supports a remote HTTP server with a custom header (Cursor, Cline, VS Code, etc.) works the same way: point it at the URL above with an Authorization: Bearer YOUR_API_KEY header. A gateway or scanner that can only forward the raw key value (no Bearer prefix) also works — both forms authenticate. Tool discovery (initialize/tools/list) doesn't require a key at all; only calling normalize_catalog does.

See docs/quickstart.md for a full copy-paste walkthrough (getting a key, per-client configs, confirming the connection with curl) if you'd rather follow one linear guide.

npm client

Prefer a typed function over hand-rolling MCP JSON-RPC calls? @acjlabs/catalog-attribute-normalizer-client (live on npm) wraps the normalize_catalog tool call:

npm install @acjlabs/catalog-attribute-normalizer-client

import { createCatalogNormalizerClient } from "@acjlabs/catalog-attribute-normalizer-client";

const client = createCatalogNormalizerClient({
  baseUrl: "https://catalog-normalizer.acjlabs.com",
  apiKey: "...",
});
const results = await client.normalizeCatalog(products, ["google", "shopify"]);

Why not just use a category classifier?

Category-classification APIs tell you what a product is. They don't touch the messier problem: standardizing attributes across sources that each spell them differently. The vendors that do take on the broader job are enterprise sales-led — demo request, annual contract, no self-serve signup and no public price. See the full comparison, or read why "clean product data" is actually two different problems (canonicalization vs. classification, and why they fail differently).

Source availability & support

This repository hosts the documentation for the hosted service. The service implementation is not open source. Bug reports and feature requests are welcome in this repo's Issues; you can also reach us at contact@acjlabs.com.