Catalog Attribute Normalizer MCP server

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

attributes (canonicalized sizes/colors/units) plus category mappings for whichever taxonomies you request (Google Product Taxonomy, Shopify, Amazon).

• Accepts a batch of raw products (title, description, raw_attributes), returns normalized

• 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, sleevelength — unrecognized keys are dropped, not passed through under a model-chosen name). Each value's provenance is "canonicalized" when it came from your own rawattributes 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 leafid 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 leafid 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.

Shortened. The full README is on GitHub.