You have the job sitting in a client's backlog: refresh the SEO title and meta description on every one of these 800 catalog pages, in their brand voice, by Friday. By hand it is a week of one-by-one edits. Raw AI is fast, but it writes straight to a site you do not own, and "I'm 80% sure I know what I just did" is not a sentence you can say to a client.
Scratch downloads the client's whole catalog as files on your laptop. Your agent reads and edits all 800 pages there, about 10x faster than it works over an API, while their live site never moves. Every change comes back as a word-level diff your account lead reviews. Nothing ships until someone approves it, and every published record is reversible.
Without Scratch, and with
| the job | without scratch | with scratch |
|---|---|---|
| Refresh SEO across a client's 800-page catalog | AI writes straight to the live site, or a junior edits pages one at a time for a week | AI rewrites all 800 as local files, the lead reviews the diff, only approved pages ship |
| Rewrite 200 product descriptions in the client's brand voice | One prompt publishes 200 off-voice rewrites with no way to read them first | Try the voice on one product, run it across 200, approve the ones that land |
| Push a client rebrand across every page | A find-and-replace catches the old name and breaks the three places it should not | The agent handles the messy cases, you see every swap as a word-level diff before it ships |
| Run a monthly content pass across every client | You rebuild the same risky workflow from scratch on each account | The same reviewed pull-edit-approve loop, repeated per client, on whatever platform they use |
| Hand the client proof of the work | A "trust me, it's done" and a screenshot | A reviewable diff and a per-record history of exactly what changed |
How it works
- Scratch pulls the client's content into files. Every product, post, page, or record from the client's Shopify, WordPress, Webflow, or HubSpot comes down to a folder on your laptop, one file per record. Their live site never moves.
- Your agent rewrites the fields you point it at. Open the folder in Claude, Claude Code, Cursor, or any agent that edits local files. Try a prompt on one record, then let it run across the whole catalog, every record, not a sample. The agent edits the files, never the client's live site.
- The lead reviews every diff and publishes. In the Scratch desktop app, each change shows next to the original, word by word. Approve what ships, and Scratch publishes only the records you approved back through each platform's API.
You can start 200 edits, go do something else, and review them all when you come back.
The economics of agency work change when the edit stops being one-by-one. The catalog pass that used to be a week of clicking becomes an afternoon: kick off the run, do other work, come back and review the diffs. One customer who builds landing pages this way puts his rate at "I can make 10 of these in an hour." And review scales the way hand-editing never did. Reviewing 500 changes takes about the same effort as reviewing 50, because you are scanning word-level diffs, not opening pages.
The deliverable improves with the margin. Instead of a status update, the client gets a record of every change next to what it replaced, and you approve the work before any of it touches their site.
What agencies use it for
Most jobs land in a client backlog because doing them by hand means opening every record one at a time, and doing them with raw AI means publishing blind to someone else's site.
- Refresh SEO titles and meta descriptions across a client's 800-page catalog.
- Rewrite a client's 200 product descriptions in their brand voice, in one reviewed pass.
- Push a client rebrand through every page: swap the old name in body copy, headings, and SEO fields at once.
- Run a monthly content pass across every client on your roster, the same loop on whatever platform they run.
- Add or fix alt text and internal links across a content-heavy client site.
- Show the client a reviewable diff and a per-record history of what changed, instead of a status update.
Why not let the AI write straight to your client's site?
An MCP server or a direct API write hands the agent the publish button straight to the client's live CMS, and it is slower anyway: every record is an API call, and on an 800-page catalog it crawls. No diff, no review queue, no rule layer, no rollback. One agency owner described running ordinary automations on client accounts as "you hope for the best and hit go and see stuff happen." One confident pass ships across the whole account, and by the time anyone on your team spots a problem it is already live on a site you do not own, with the original often gone. On a client account, that is not a bug to fix later. That is the relationship.
Scratch gives the agent the same full read-and-write access, but against a local copy. Scratch pulls the publish step out and hands it to your account lead. The agent can change anything; only a person can ship it, and only after reading the diff. Every published record is reversible from Scratch, so a bad edit on a client's site is a revert, not an apology email.
What's safe, and what's off-limits
You bring your own AI. Scratch holds no AI credentials and runs no model, so there is no per-seat AI bill stacked on every client account. You sign into Claude, Claude Code, Codex, Cursor, Copilot, Cline, or Windsurf the way you already do. By default, nothing leaves your laptop until the lead publishes, and every published change is reversible per record. For jobs you run on every client, optional Python validators flag bad edits next to the diff, like an over-length SEO field or a field you have marked never-touch, so the mechanical mistakes get caught early and review time goes to the calls that need a human.
Every client's data is also backed by a git repository from the first sync, with full version history. One agency owner's verdict on that: "It is not the sexiest sounding thing, but it is a real lifesaver. The one time you need it, it's like insurance. We all hate paying for it, but it sure is nice when you break your arm."
Questions agencies ask
Can I run AI across a client's whole catalog without risking their live site?
Yes. The agent reads and rewrites the whole catalog as files on your laptop, not on the client's site. Nothing reaches their live CMS until your lead has seen the change as a diff and approved it. The risky step, publishing, is the one step Scratch keeps in human hands.
How much client work can one person actually run?
More than the hours suggest, because the person stops doing the edits. You can start 200 edits, go do something else, and review them all when you come back. Reviewing 500 changes takes about the same effort as reviewing 50. One customer puts landing pages at ten an hour. The constraint becomes review judgment, not clicking speed.
What do I show the client so they trust the work?
A reviewable diff and a per-record history. Scratch records every change next to the original, word by word, and keeps a per-record history of what shipped. You can walk the client through exactly what changed and when, instead of asking them to take your word for it.
Can I undo a single bad change on a client account?
Yes. Every published change is reversible per record, even on platforms with no native undo. A mistake on one product or one page is one record to revert, not a full restore of the account.
Does this work across all my clients' platforms?
Yes. Scratch pulls from Shopify, WordPress, Webflow, HubSpot, and more, into the same local files and the same review loop. You run one workflow across a roster of clients on different platforms.
Will this add a per-seat AI cost to every client?
No. Scratch runs no model and holds no AI credentials. You bring the agent you already pay for and sign in the way you already do, so there is no AI bill stacked on top of each client account.
How do I keep junior work from publishing before a lead sees it?
Publishing is a separate, human step. A junior can run the agent across a client's whole catalog and never touch the live site, because only the diff approval ships anything. The account lead reviews and approves, so nothing reaches a client without a second set of eyes.
See it on your own client sites
The fastest way to trust it is to watch it run on a real client catalog. See it run on your client's content, where Curtis connects a platform live and shows your own content as a reviewable diff, no pitch and no slides. Or download Scratch free and run the first pass yourself.