The right level of automation for an AI agent is 99%. The right level of access is read and write everything. The remaining 1% is review and publish. That last 1% stays human.
Scratch is the loop that makes this work against your real content. You connect Scratch to wherever your content lives. Shopify, WordPress, Notion, HubSpot, the others. It downloads everything as a folder of files on your laptop. Claude has full read and write access there. You tell Claude what you want done. Rewrite the meta descriptions on these 400 posts so each one is under 158 characters and ends with a call to action. Claude does it.
Then Scratch shows you every change as a word-level diff against the original. You do the 1%: approve what ships. Scratch publishes only those records back to your CMS, one at a time. The records you didn't approve never leave your laptop.
The question you're actually asking is whether you can trust Claude with your live site. You can. Publishing is the thing Claude doesn't get to do.
Where you drive Claude from
The Claude desktop app has three tabs across the top: Chat, Cowork, and Code. They all read and write the same files in your Scratch folder, so you can move between them as the job changes.
Chat is what you'd open first. Drag a Scratch file into the window, tell Claude what you want, save what comes back over the original. Best for one or two records, or for working out whether your prompt is any good before you point it at the whole catalog.
Cowork is what you reach for when the job is bigger than chat. Point Cowork at the Scratch folder, describe the outcome you want, and walk away. It plans the work, edits every file that needs editing, and comes back to you when it's done. Same agent as Claude Code, just running inside the desktop app instead of the terminal.
Code is Claude Code, embedded in the desktop app instead of the terminal. Same loop, same skills, same diff surface in Scratch. If you already live in a terminal, the Claude Code page is the one you actually want.
You bring your own Claude. Scratch holds no Anthropic credentials and runs no model. You sign into the Claude desktop app the way you already do. Pro, Max, Team, whatever you're on, it works.
The things you're probably worried about
Will Claude publish anything to my live site? No. Claude only reads and writes files in your Scratch folder. Scratch is the only thing that talks to your CMS, and it only publishes what you've clicked approve on, one record at a time.
Can I undo if something ships I didn't want? Yes. Every published record is reversible from the Scratch app, per row. The original sits next to the rewritten one until you tell Scratch which one stays.
What about my prices, SKUs, member data, anything sensitive? You lock those fields when you set up the connector, and Scratch can't write them back even if Claude tries. On top of that, you can run Python validators that reject any diff that touches a field you've flagged. Bad rewrites get filtered out before you have to look at them.
Does Anthropic see my content differently because it came through Scratch? No. The data path is your laptop, to your Claude account, back to your laptop. Whatever your Claude settings say about training and history applies. Same as if you typed every word in by hand.
How is this different from an MCP server or a Claude plugin for my CMS? Both give Claude the publish button. Scratch doesn't. Claude gets the same full access an MCP would give it, but the publish step stays in your hands. That's the whole difference.
Can I do this on hundreds of records at once? Yes, that's the use case. Cowork and Code are built for catalog-scale jobs.
Do I need to be technical? No. You install Scratch, connect a source, point Claude at the folder, and click approve on the diffs. Validators are optional and only matter once you're running the same job often enough to want guardrails.
See Claude connected to each platform
Browse the skills below for prompts that work end-to-end with Claude.