Use Hermes with Scratch

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.

Hermes is an open-source CLI agent by Nous Research. It runs persistently, with memory across sessions, and it builds reusable skills from experience. That changes the shape of a Scratch workflow in a specific way: the first time you run Hermes on a Scratch folder, it learns your catalog. The second time, it starts from that skill instead of starting from scratch. The job runs faster because the context is already there.

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. You open a terminal, cd into the Scratch project folder, and tell Hermes 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. Hermes does it, writes the files, and builds a skill it can reuse the next time you run the same job.

Then Scratch shows you every change as a per-record diff, the way you'd review a pull request. You do the 1%: approve what ships. Scratch publishes only those records back to your CMS or CRM, one at a time. Hermes edits. You keep the merge button.

How to run Hermes on a Scratch folder

1. Pull your records into Scratch.
2. cd into the Scratch project folder in your terminal.
3. Run hermes to start a session.
4. Give Hermes a prompt describing what you want done across the catalog.
5. Hermes uses its built-in tools (web search, browser automation, vision, file access) to plan and execute the work. It edits files in place.
6. When it finishes, open Scratch. Every changed record shows a word-level diff against the original.
7. Review per record. Approve. Publish.

Hermes supports subagents, so you can split large catalogs into parallel threads if you want to run multiple passes at the same time. The files land in the same Scratch folder either way, and the diff surface handles them the same.

What the memory means in practice

Most agents start cold every session. You re-explain the job, the field names, the tone rules, the length caps. With Hermes, that overhead compresses after the first run.

When Hermes finishes a Scratch session, it can save a skill: a reusable set of instructions that describes your catalog structure, your prompt, and the decisions it made along the way. The next session loads that skill and skips the orientation phase. If you run the same job weekly, by the third week Hermes is fast enough to schedule.

Hermes supports scheduled automations natively. You can wire a Scratch job as a recurring task: pull updated records, run the skill, write the files. Scratch still holds the review step. Nothing publishes until you say yes, but the preparation work happens on a schedule instead of on demand.

Subagents extend this further. A coordinator agent can split a large catalog by category, hand each slice to a worker agent with the relevant skill loaded, and merge the results back into the Scratch folder. Scratch sees a unified diff either way.

What stays safe

Hermes has full read and write access to your Scratch folder. That is exactly as much access as it needs, and no more.

Scratch is the only thing that talks to your CMS. Hermes edits files. It cannot publish. It cannot touch your live site. The publish gate stays in the Scratch app, one record at a time, with your approval on each one.

If you want an additional layer, you can run Python validators in the Scratch review step. They run before any record reaches the approve queue. A validator can reject any diff that touches a locked field, exceeds a length cap, or fails a tone check. By the time you're reviewing, the bad rewrites are already gone.

You bring your own AI account and tokens. Scratch holds no Hermes credentials and runs no model. You sign into Hermes the way you already do.

Browse the skills below for prompts that work end-to-end with Hermes.