Catalog Attribute Normalizer

What Catalog Attribute Normalizer does

The Catalog Attribute Normalizer MCP server processes batches of raw products and produces structured results with normalized attributes and category paths. Each product can include a title, description, and raw_attributes object. The response includes a schema version, the source title, category mappings, and normalized attribute values.

Supported controlled attribute keys include size, color, material, gender, and sleeve_length. Unknown keys are discarded rather than retained under names selected by a model. Values already supplied in raw_attributes are cleaned deterministically where possible. Values inferred from the title or description are marked as extracted, so downstream systems can review them as suggestions.

Category targets can include Google Product Taxonomy, Shopify, and Amazon. Google and Shopify candidates are retrieved from taxonomy files and checked before a leaf ID is returned. Amazon classification is best-effort because the service does not have a comprehensive public Amazon taxonomy reference file.

How it works

Call the normalize_catalog MCP tool with a products array and a target_taxonomies array. A batch may contain multiple products, but usage is counted per product rather than per request. The service can return category paths, leaf IDs where independently verified, confidence values, and attribute provenance.

Canonicalization and classification have different reliability characteristics. Existing input attributes are normalized without trying to discover missing values. Inferred attributes and category results should be reviewed when marked as extracted, when confidence is low, or when a category has a null leaf ID. In the README's 12-product evaluation, Google and Shopify produced 22 exact path-and-leaf-ID matches out of 24 checks.

Setup and configuration

The hosted MCP endpoint is https://catalog-normalizer.acjlabs.com/mcp. An alternate worker URL points to the same deployment. Obtain a free-tier key by sending an email to POST /v1/signup; the returned key is shown once and cannot be recovered. The free tier allows 500 products per month without a card.

Configure an MCP client that supports a remote HTTP server and custom headers with an Authorization: Bearer YOUR_API_KEY header. Claude Code can be configured with claude mcp add --transport http. The repository also documents a typed npm client, @acjlabs/catalog-attribute-normalizer-client, for applications that prefer a function call over direct MCP JSON-RPC handling.

Tool discovery does not require a key, but invoking normalize_catalog does. Batches exceeding the remaining allowance are rejected in full with HTTP 402, and failed calls do not consume units.

Tools and capabilities

The Catalog Attribute Normalizer MCP server exposes one documented MCP tool:

• normalize_catalog: accepts product records and requested Google, Shopify, or Amazon taxonomies, then returns normalized attributes and category mappings.

The same normalization operation is available through a plain HTTPS API. Results include provenance for attributes, confidence for category paths, and verified leaf IDs where applicable. The service is intended for catalog cleanup and feed preparation rather than unrestricted attribute extraction.

Limitations and notes

The implementation is hosted and is not open source; this repository contains its documentation. Amazon mappings are not retrieval-grounded in the same way as Google and Shopify mappings. Confidence values and null leaf IDs require review rather than automatic acceptance.

The free allowance is measured per product, including products inside a batch. Paid plans are described as forthcoming in the README. A lost key cannot be recovered through the documented signup flow; a new signup is required.