This is the first of five open-source projects from MACH Alliance Agent Solution Studio, with the other four to follow in the coming weeks.
Every commerce demo needs a product catalog. Teams usually reach for one of three options. They reuse a stale sample dataset, scrape real products, or build a spreadsheet by hand. Sample data looks generic in front of a client. Scraped data carries real brand names and images you have no right to publish. Hand-built data takes days.
Then the format problem starts. A search platform, a commerce backend and a merchandising tool each expect product data in a different shape. Someone has to reshape the same catalog for each one, and the work starts over when the demo changes.
The Product Catalog Generator handles both steps. It creates realistic products with AI-generated copy and images, then exports them in the exact format a target system expects. It runs locally and is released under the MIT license.
Say a solutions engineer is preparing a BigCommerce demo for a home goods retailer.
She creates a catalog called Home, adds a Vases category and writes a base image prompt that sets the photo style for everything in it. Then she types one line: “12 minimalist ceramic vases, $40 to $120.” The text model writes names, descriptions and prices for twelve products. The image model generates a photo for each one from the catalog’s base prompt plus the product description.
One vase comes back with the wrong glaze. She edits that image in place instead of generating a new one. Every generated image also passes through automated pixel correction, so backgrounds come out pure white. Models tend to render “white” a few shades off, which shows up the moment products sit side by side on a storefront.
None of the products carry real brand names. Models asked for realistic products tend to borrow real companies, which has no place in published demo data. Brand names are off by default, and an optional environment setting invents a fictional house brand instead.
Next she opens the built-in BigCommerce export template, previews the mapping against one real product, and downloads the result as JSON plus a zip of images. The template already knows details that trip people up, such as BigCommerce storing custom field content under text rather than value.
If her next demo runs on a different search platform, she keeps the catalog and builds a new export map. For simpler targets, an AI-assisted flow drafts the map for her.
The AI drafts the map. Code runs the export. Once an export map exists, running it makes no model calls, so every export is fast and gives the same result each time. Adding a new built-in template takes one entry in the code.
The tool is headless and API-first. The web UI and the catalog API read the same data, and the UI has no access to it that the API doesn’t also expose. Another service can use the catalog as a backend, reading products and managing images through a REST API protected by an optional token.
It is also model-agnostic. You choose the text, fast text and image models separately in Settings, and reach them directly through OpenAI or Anthropic, or through Vercel AI Gateway, OpenRouter or any OpenAI-compatible endpoint.
Fork it if you build commerce demos, proofs of concept or integration tests and need data that looks real. System integrators can use it to stand up client demos faster. Vendors can add an export template for their own platform so demo data arrives ready to load.
The project shows that AI can do the creative work in a commerce pipeline while deterministic code handles the parts that have to be exact.
Each piece of the stack has a specific job. The Vercel AI SDK keeps the model layer swappable, so no team is locked into one provider. Anthropic models handle text, including schema lookup by web search when drafting an export map. OpenAI’s image model is the default for product photos because it can both generate and edit images, and Anthropic publishes no image models. BigCommerce, Coveo and Algolia are the export targets, chosen because their data structures are where hand-reshaping takes the most time. The catalog itself is stored in Algolia’s hits format.
The Catalog Generator is a prototype for demos and integrations. It has no checkout, payment, inventory tracking or customer account features. Products are stored as flat JSON files on disk, which suits demo-scale catalogs. The routes the local web UI uses have no authentication, and paid model calls, including image generation, have no built-in spend tracking. Run it on your own machine in Docker or another disposable sandbox, and keep it off public hosts and shared networks.

The code is public on GitHub under the MIT license. The getting started guide covers requirements, credentials, safe setup and your first catalog and export.
Contributions are welcome. Fork the repo, add an export template for a system you work with, or open a pull request with an improvement. The contributing guide explains how to run the tests and what a change needs to include.