Connect Claude to your product catalog with MCP

Most catalog work isn’t hard, it’s wide: the same missing attribute on four hundred products, a GTIN that never made it onto half the variants, a channel that won’t publish until three fields are filled in everywhere. The minipim MCP server gives an AI agent the tools to do that work against your real catalog — reading the schema first, fixing one attribute at a time, and leaving a full audit trail.

MCP is the open protocol Claude and other assistants use to call external tools. The minipim server is published on npm as @minipim/mcp and runs locally next to your assistant.

What an agent can do

The server exposes ten tools, built around minipim’s channel readiness model:

Tool	What it does
`get_catalog_overview`	Counts, families, channels and the attributes each requires, locales, categories, brands
`get_family_schema`	Every attribute a product family accepts, its value shape, and whether it’s required
`search_products`	Text and filter search
`get_product`	One product in full — attributes, media, variants, identifiers
`find_incomplete_products`	Products scored against each channel’s requirements, worst first
`get_product_completeness`	One product’s readiness gaps, per channel
`set_product_attributes`	Set or clear attributes on a product (write)
`bulk_update_products`	One change across up to 100 listed products — dry run by default (write)
`set_variant_identifiers`	UPC, GTIN, MPN or SKU on a variant, with GTIN check-digit validation (write)
`check_connection`	Confirms which organization the key belongs to

Writes are guarded: set_product_attributes refuses attribute codes the product’s family doesn’t declare, and bulk_update_products takes an explicit list of product ids rather than a filter, so an agent can’t accidentally rewrite the whole catalog.

Three built-in workflows

The server also ships three prompts your assistant can start from:

audit-catalog — read-only. Reports what’s incomplete and why, without changing anything.
enrich-catalog — works through gaps one attribute at a time across many products, which is how catalog cleanup actually goes fastest.
prep-for-channel — takes one channel and gets products to the point where they can publish there.

Every prompt instructs the agent never to invent a measurement, a barcode, or a compliance claim to fill a gap — it reports the gap instead.

Setup

1. Create an API key. In minipim, open System → API keys in the sidebar and create a key. The key carries its organization, so there’s nothing else to configure. If you only want audits, set MINIPIM_READ_ONLY=true (below) and the write tools disappear.

2. Add the server to Claude Code:

claude mcp add minipim --env MINIPIM_API_KEY=pim_your_key -- npx -y @minipim/mcp

Or to Claude Desktop (and other clients that use the same config format), in the MCP servers config:

{
  "mcpServers": {
    "minipim": {
      "command": "npx",
      "args": ["-y", "@minipim/mcp"],
      "env": { "MINIPIM_API_KEY": "pim_your_key" }
    }
  }
}

3. Ask. “Run the audit-catalog prompt” is a good first message. check_connection confirms you’re pointed at the right organization.

Configuration

Variable	Purpose
`MINIPIM_API_KEY`	Required. Your API key.
`MINIPIM_BASE_URL`	Defaults to `https://api.minipim.com`. Self-hosted instances point this at their own API.
`MINIPIM_READ_ONLY`	`true` hides the three write tools and offers only the read-only audit prompt.

The package also installs minipim-mcp-http, a streamable-HTTP version for clients that can’t launch a local process. It runs as a single tenant using the key in its own environment, so run it somewhere private — don’t expose it to the internet.

Limits today

Products and variants only. Content pages, media uploads, categories, brands and modifiers aren’t exposed to agents yet.
No hosted endpoint. The server runs on your machine (or your own server); there’s no one-click remote connector for claude.ai yet.
Deliberately excluded: deleting products, wiping the catalog, managing members or API keys, and starting connector syncs. Those stay in the admin.

Every write an agent makes goes through the same REST API as the admin, so it shows up in the audit log and fires the same webhooks — a storefront subscribed to product changes picks up an agent’s edit exactly as it would a person’s.