A tone pass across 3,000 products is not 3,000 small jobs. It is one job, and it only comes out consistent if the agent can hold the whole catalog while it works. Grok Build can. Its 2 million token context takes a large slice of your products, plus your rules and conventions, in a single pass, and its plan mode writes out what it intends to do, file by file, before it touches anything. What has stopped you is not the agent. It is the storefront: live, taking orders, and wired so that most AI tools write straight to it.
Scratch unwires that. It pulls your catalog down as a folder of files on your laptop, one per product, and Grok Build does 99% of the work there, in the copy. The last 1% is yours: every change comes back as a per-record diff, you approve what ships, and Scratch publishes only those records back to Shopify. Prices, variants, and inventory stay locked the whole way through.
How it works
- Scratch pulls your catalog into files. Products, articles, blogs, and pages land in a folder on your laptop, one file per record, with an
AGENTS.mddescribing the layout, which Grok Build picks up automatically. - Grok Build plans, then edits.
cdinto the folder, rungrok, and launch in plan mode. It writes out the pass file by file: which products, which fields, what changes. You read the plan, approve it, and Grok Build works through the catalog, holding your rules and a large slice of records in context the whole time. It edits the files, never the live store. - You review every diff and publish. Scratch shows each change beside the original, word by word. Approve what ships, and Scratch sends only those records back to Shopify. What you skip never leaves your laptop.
What people use it for
The catalog work that needs the whole catalog in view:
- Run one tone pass over every product description, so the store reads like one writer.
- Dedupe supplier boilerplate that fifty products share word for word.
- Backfill SEO titles and meta descriptions, each unique, each under the cap.
- Normalize vendor, type, and tags so collections and filters behave.
- Carry a rename through products, articles, and pages in a single pass.
Plan mode on a dozen products first. When the plan reads right, run the catalog.
Why not an MCP server?
A Shopify MCP server wires the agent straight to your live store, with the write bundled into the same connection as the read. One confident pass lands on every product page your customers are looking at, and you review it after the fact, if at all.
Scratch gives Grok Build the same full read and write against a local copy. The publish step is lifted out of the agent loop and handed to you, per record. Grok Build can change anything in the files. Only you can ship it.
What Grok Build edits in Shopify
Editable fields:
- Product titles, descriptions (HTML), and handles
- Vendor, type, and tags
- Product SEO title and meta description
- Articles, blogs, and pages
Stays locked:
Prices, variants, inventory, and metafields are locked at the connector level. Grok Build cannot write them back even if it tries, and length and taxonomy validators flag problems right on the diff. The full field list lives on Scratch for Shopify.
Questions people ask
Is this an MCP server or an xAI integration?
Neither. An MCP server hands the agent a live line to your store with the write built in. With Scratch, Grok Build works on local files, and publishing is a separate step you approve, one record at a time.
Can it touch my prices or inventory?
No. Prices, variants, inventory, and metafields are locked at the connector level, so a stray edit to them cannot write back. The pass you cannot afford to get wrong is the one the connector refuses.
What does plan mode add on top of the Scratch review?
A second gate, in front. Plan mode shows you what Grok Build intends before it edits a single file, and the Scratch diff shows you what it actually did before anything ships. A bad instruction gets caught as a plan, not as three thousand rewrites.
Can I run this on a schedule?
The editing half, yes. grok -p runs headless, so a weekly pass over new arrivals can prepare its edits on a cron. The publish half never runs itself: the diffs wait in Scratch until you approve them.
Can I roll back a published change?
Yes. Every published record is reversible from Scratch, per row. The original sits next to the rewrite until you decide which one stays.
Do I need to be a developer?
You need to be comfortable in a terminal, since that is where Grok Build lives. If you would rather have the same loop in a desktop app, the Claude desktop app runs it with a more familiar surface.
See it on your own catalog
The fastest way to trust it is to watch it run on your products. Book a 30-minute demo on your Shopify catalog →, or try Scratch free and run the first pass yourself.