Keepp's AI Editor vs Handing an Agent Your API Key: Which One to Use When
Two ways to change a Keepp page with AI: the assistant inside the editor, and an outside assistant holding an API key. They behave differently on undo, on what they can see, and on when your visitors notice.
Keepp has two ways to change a page with AI, and they are not versions of each other. One is a panel inside the editor. The other is an API key you hand to ChatGPT, Claude, or Gemini along with the Keepp skill, so your page becomes something that assistant can manage.
Both are Pro. Both write the same blocks. The differences worth choosing on are what each one can see, what happens the moment it is wrong, and whether your visitors are watching while you work.
What each one actually is
The editor panel sits in the page editor: a tab beside Add and Page style on a computer, a button next to Add on a phone. You describe a change, it edits the page in front of you, and you look at the result.
The agent route is an API key, generated under AI Agent in your dashboard, plus a skill file that teaches an assistant how a Keepp page is built. You then work in that assistant, wherever you already work, and your page is one of the things it can do.
The difference that matters most: when it goes live
The editor panel edits your draft. Your published page keeps showing what it showed until you press Publish. If an edit is wrong, you undo it and nobody saw anything.
An agent write is a publish. The API replaces your page and that is the live page, immediately. There is no draft step and there is no undo, which is why the skill tells an assistant to read the page before writing and to send the whole thing back.
Neither behaviour is the better one. A published write is what makes the agent route useful for anything scheduled or scripted, where nobody is sitting there to press a button. A draft is what makes the editor panel safe to experiment in on a Saturday morning.
What each one can see
The editor panel sees the page as it is on your screen, including work you have not published. That is the version you are asking about, so that is the version it edits.
An agent reads what you last published. If you have unsaved edits open in a tab somewhere, it cannot see them, and a write from the agent while your editor is open on older work is the kind of collision the version check exists to catch.
What happens to the blocks you did not mention
In the editor panel, a block the assistant does not mention stays exactly where it is. That is deliberate: a model answering about one heading on a long page will often return only the parts it touched, and forgetting has to be survivable.
Through the API, leaving a block out means deleting it. That is a reasonable contract for a whole-document write, and it is why the skill’s loop is read, modify, write the whole thing back rather than send the changed part.
What neither of them will do
Both work from what your account already holds. Neither invents a price: a product block points at a product record, and the title and price come from that record, so the page shows what you actually charge. Neither can point a checkout at something that does not exist, and neither will write an address or opening hours you have not given it.
The agent route can run without you
This is the capability the editor panel does not have, and it is the main reason to set up a key.
The API is three plain HTTP endpoints: read the page, replace the page, list the products and bookings it can point at. Anything that can make an HTTP request can drive it, on whatever schedule you like, with nobody watching:
- A weekly refresh. A script reads your page every Monday morning, swaps the three featured cards for this week’s products, and writes it back. Your page is current before you open your laptop.
- Sold out, off the page. Your stock system drops to zero and the job that notices it removes that product block on the same run, instead of at the end of the day when you remember.
- Seasonal swaps. The December section goes up on the first and comes down on the second of January, from a job you wrote once.
- Several pages at once. If you run more than one page, a script sets up all of them from the same source instead of you repeating yourself in an editor.
- Inside an assistant you already use. No schedule at all: your page is simply one of the things the assistant you work in can change, in the same conversation as everything else.
Rate limits are 60 requests a minute, which is a long way above what a scheduled job needs and worth knowing before you write a loop.
None of that is reachable from the editor panel, which needs you in the editor, on the page, in the moment.
Which to use
Use the editor panel when you are looking at your page and want it different: the class moved, the heading is too long, the two cards should sit side by side. It is faster for one change, it needs nothing set up, it sees work you have not published, and you can undo your way out of anything. Its limit is you: it changes the page in front of you, while you are there.
Use the agent route when the change is part of something repeatable or unattended, or part of a bigger job in an assistant you already work in. Its limit is that it publishes as it writes, so the care has to live in the read, modify, write loop rather than in an undo button, and it only ever sees what you last published.
A rough rule: if you would describe the task as “change this”, the panel is quicker. If you would describe it as “keep this up to date”, it wants the agent.
Plenty of people will use both, and they do not conflict. The agent writes the published page, the editor edits the draft, and the version check keeps them from overwriting each other silently.
Getting started with either
For the panel, open your page in the editor and tap AI. It is free for the first 24 hours after you create your page, then part of Pro. Steps are in Edit your page by asking.
For the agent, generate a key under AI Agent and give your assistant the skill. Steps are in Manage your page with an AI agent.