Help · Settings

Connect your MCP AI assistant

You can connect your own AI assistant to your store. Once it is connected, you can ask it about your store and ask it to write things for you — an email, a campaign, an automation, a segment — and it puts them in the app as drafts, where you open them, change what you want, and decide what happens next.

It works with Claude, ChatGPT, and any other assistant that speaks the open standard called MCP — look for connectors, integrations, or custom tools in its settings.

What it never does

Your assistant cannot send anything. It writes drafts, and you decide what happens to them: it drafts, you send.

It cannot send an email or a text, schedule a campaign, enable a message inside an automation, publish a form, enable an automation, cancel, pause or disable anything that is already running, delete anything, export your people, change anyone's consent, or change a setting. Those are not permissions we forgot to give it. They are refused in the same place your team's own permissions are refused, before anything else is looked at, and the connection holds nothing it could use to try.

That is the whole point of it. An assistant that could press send would be a way for a mistake — yours, ours, or a stray instruction it read somewhere — to reach the people whose profiles you hold with nobody looking. An assistant that can only draft is one you can let loose on a Monday morning and still read everything before it goes out.

How to connect one

Open Settings → MCP AI assistant. That page shows an address, with a button beside it that copies the address for you.

The address is the same for every store, so it is not a secret. What decides the store is the store you are in when you allow the connection. The box that asks you names it, and your assistant works in that store and no other. If you work in more than one store, choose the store first, then connect: open the app, press your store's name at the foot of the menu, choose the store you want your assistant to work in, and only then add the address in your assistant. If the app asks Choose a store on the way, the store you choose there is the one the box names.

Then, in your assistant, add that address as an MCP connector. Every assistant words this differently — look for connectors, integrations, or custom tools — but they all end up asking you for an address, and that is the one to paste.

Your assistant will then send you here to sign in. Sign in the way you normally do, and a box opens over the app asking you to allow the connection — read what it says you are about to allow, and press Allow. That is the moment the connection is made, and nothing reaches your store before it.

That box names the assistant that is asking, under It says it is, and the address it goes back to, under Its address. The name is whatever the assistant calls itself — anybody can pick a name — so the address is the part worth reading. If you do not recognize it, press Don't allow.

You are signing in as yourself, so the connection belongs to you and to this store. If you work in more than one store, add Carpe Messaging to your assistant once for each store, and allow each one from the store it is for. If you try to allow the same connection in a second store, the box says:

This assistant is already connected to another store you work in. To connect this store too, add Carpe Messaging to your assistant again and approve it here.

Nothing changes in the first store when that happens.

What it can read, and what it can write

Your assistant can read your campaigns, your automations, your segments and what they count, your lists, your email templates, the images in your Images & media, your brand from Settings → Store, and your reports. Reports come back as totals — how many were sent, opened, clicked — never as a list of who did what.

It can write drafts: email templates, draft campaigns, draft automations and segments. A draft it writes is an ordinary draft — it shows up where your own drafts show up and you edit it the way you edit anything else. The app records that your assistant wrote it, and which team member it was acting for.

It can also add an image to your Images & media, one at a time, from a web address — an image on your own website or your Shopify store, for example. The address has to be a public https:// one, and the image is a JPEG, PNG, WebP or GIF, checked the same way as one you upload yourself and counted toward the same library. It shows up in Images & media like any other image, and the app records that your assistant added it. A viewer's assistant cannot add one. Your assistant is asked to say what each image shows as it adds it — that is the image's description — because, as in the app, an image with no description cannot go in an email until someone describes it.

An email your assistant writes as HTML sends as it was designed. Its style rules — how it stacks on mobile, larger headings on a small screen, links without an underline — its background colors and its class names are kept, fenced in so they can only change that email's own part and never your footer or unsubscribe link. Two things to tell your assistant. Ask it to put its images in your Images & media first — it can look through the ones already there, or add one from a web address — and to use the address it gets back. An image or background image served straight from another website is removed, and the draft says so. And do not load fonts from another website — those are removed too, so name a font your readers already have.

Your assistant can build an email out of the same blocks you use in the editor — text, images, buttons and the rest — so the template it saves opens in the email editor block by block, and you change it the way you change one you built yourself. Give it your ideas, your images and your style guide, and it can read your brand from Settings → Store on its own: your colors, your link style and your logos. If you have not saved a brand there yet, tell your assistant your colors instead. A template it saves sits with your other templates and is never sent until you choose it for a campaign or an automation. The Design an email with your assistant page has what to give it and an example to start from.

If an automation has already been enabled, your assistant can work on its working copy and nothing else. The version people receive keeps running exactly as it is until you apply the changes yourself.

If a segment is being used by a form that is live on your store, your assistant is refused the edit and told which form. Changing that rule would change who sees the form on your storefront, which is a live thing, not a draft. It can make you a new segment instead.

Roles, and what an assistant may do

An assistant has exactly the permissions of the team member who connected it, and never more. A viewer's assistant only reads, and even an owner's assistant can only read, draft and add images.

A viewer's assistant can answer questions about the store and cannot write a draft at all. A marketer's assistant reads, drafts and adds images. A manager's or an owner's assistant reads, drafts and adds images, and that is still all — the things owners can do that marketers cannot, like managing the team or the store connection, are on the refused list for every assistant, whoever connected it.

Change somebody's role and their assistant changes with it, on its next request. Remove them from the team and their connection is revoked for good.

Disconnecting

Settings → MCP AI assistant — the same page that shows the address — lists every connection: which assistant, which team member, when it was connected, and when it was last used. Revoke ends one immediately: the next thing that assistant tries stops at the door.

Owners see every connection on the store. Everybody else sees their own.

A connection also ends on its own if nobody uses it. It lasts as long as your assistant keeps using it, and after 30 days without being used it stops working. Nothing is lost when that happens — you connect the assistant again the same way you connected it the first time. The page marks a connection that has gone quiet, so you can see it coming.

Setting a new password, or changing it, disconnects every assistant you connected, in every store — connect them again afterwards.

Revoking does not touch anything the assistant wrote. Drafts are yours; they stay where they are.