Connect an AI assistant (MCP)
Connect ChatGPT, Claude or another AI assistant to your store, then work on your blog by chatting instead of clicking through the app.
What is MCP?
MCP (Model Context Protocol) is the standard AI assistants use to talk to an outside app. Connecting Avada AI Blog Builder over MCP gives your assistant a set of tools scoped to one store — the one you connect it to. It can read your blog and, depending on how you connected it, write to it.
Once connected you can ask things like:
- "How many posts do I have, and how many are still drafts?"
- "What's going out next week?"
- "Suggest keywords for a post about winter skincare"
- "Plan an outline for that post, then write it"
- "Write a TL;DR for my last post"
- "Publish the draft you just wrote, next Monday morning"
You'll find it in the app under MCP.
Two ways to connect
Which one you use depends on the assistant — but in this app the two are not equivalent, so read the table before you pick.
| Sign in with Shopify | Access token | |
|---|---|---|
| Who it's for | ChatGPT, Claude | Editors and terminal assistants |
| How it works | You sign in through your browser and approve the connection | You paste a token into the client's config file |
| What it can do | Read and write — write, edit, publish and schedule posts | Read-only |
| Set up from | The Connect buttons on the MCP page | The Access tokens card on the MCP page |
A token created on the MCP page can only read. If you want an assistant to write or publish posts, connect it by signing in with Shopify — that is ChatGPT and Claude today.
Pick your assistant
| Assistant | Connects with | Can write? | Guide |
|---|---|---|---|
| ChatGPT | Sign-in | Yes | Connect ChatGPT |
| Claude (web, desktop, mobile) | Sign-in | Yes | Connect Claude |
| Claude Code | Token | No | Connect Claude Code |
| Cursor | Token | No | Connect Cursor |
| Gemini CLI | Token | No | Connect Gemini CLI |
| Codex CLI | Token | No | Connect Codex CLI |
| Anything else | Token | No | Other clients |
Where to find your endpoint
Every token-based guide starts here.
Open the MCP page
In the Avada AI Blog Builder app, click MCP in the left menu.
Copy the endpoint
Scroll to Setup guide and copy the MCP endpoint at the top of the card:
https://…/mcpThe endpoint is the same for every client and every store. What ties a connection to your store is the token you pair it with, or the Shopify account you sign in with — not the URL. So a token is the thing to keep private, and a token is the thing to name carefully if you run several stores.
Access tokens
ChatGPT and Claude sign in with your Shopify account. Editors and terminal assistants can't — they have no browser to send you to — so they use an access token instead: a long string you paste into the client's config once.
Which assistants need one
A token is required for — Claude Code, Cursor, Gemini CLI, Codex CLI, and any other MCP client.
No token needed for — ChatGPT and Claude. They sign in with your Shopify account instead, and a token would give them less than signing in does.
A token is read-only. A client holding it can read your posts, your schedule, your plan and your AI token balance, but it cannot write, edit, publish or schedule anything. Writing needs a connection you signed in for — ChatGPT or Claude.
Read-only is still worth protecting: a token reaches your store's content, and anyone holding it reaches the same content.
Create an access token
Click Create token
On the MCP page, find the Access tokens card and click Create token.
Name it
Give it a Name you'll recognise later — the name is how you find the right token to revoke, so Claude Code on my laptop beats token 1. A Description is optional and shows in the list under the name.
Copy the token
The token appears once. Copy it now and paste it into your client's config, following the guide for that client from the table above.
Once you close this dialog the token is never shown again — the Done button only becomes available after you've copied it. If you lose it, revoke it and create a new one; there is no way to look it up.
Manage your tokens
Where tokens live
The Access tokens card lists every token you've created, with its Name, a masked Value so you can tell them apart, its Access level and when it was Created. Use the search box if the list gets long — it matches on the name, the description and the masked value.
One token, any client
A token isn't tied to the client you created it for: the app checks the token, not who is presenting it. The same token works in Claude Code, Cursor, Gemini CLI and anything else.
That cuts both ways. Revoking a token cuts off everywhere you pasted it, so one token per client per machine is the tidier habit.
Revoke a token
Click the delete icon on the token's row and confirm. Any client using it loses access to your store straight away, and revoking cannot be undone.
Revoke a token when:
- you no longer use the client it was for
- it was on a computer you no longer have
- you pasted it somewhere it shouldn't be — a shared config file, a screenshot, a chat message
Keeping a token safe
- One token per client, per machine. Sharing one token across several places means revoking it cuts off all of them.
- Don't commit it. A token in a config file in a git repository is a token everyone with repo access has.
- Don't paste it into a chat. Paste it into the client's config file, not into a conversation with the assistant.
What your assistant can do
Reading is always allowed. Writing is only there on a connection you signed in for.
| What the assistant can reach | |
|---|---|
| Always allowed — reading | Your store domain, plan and post language · your list of posts with status and last change · an exact post count · one post in full, including its SEO score and failing checks · what's scheduled next · your AI token balance · keyword suggestions for a topic · the app's own help pages |
| Writing — sign-in connections only | Plan an outline · write a new post as a draft · save a post you wrote yourself · rewrite a post for SEO · write meta title, meta description and featured-image alt text · write a reader-facing TL;DR · write alt text for images inside a post · set any field to your exact wording · publish, schedule or unpublish |
Nothing goes live by accident
Three rules the assistant works under, all built into the app rather than left to the assistant's good manners:
- Writing lands as a draft. A post the assistant writes is saved unpublished. You review it in the app.
- Rewrites land as a draft version. Asking it to rewrite a post, write metadata or write a TL;DR saves a draft version for you to approve — the live post is untouched until you approve it.
- Publishing is a separate, explicit step. Publishing is the only thing a shopper ever sees the result of, and the assistant is instructed to ask you first, and to ask whether you mean now or a set time.
Two things do write straight to the post: setting a field to your exact wording ("change the title to X"), and writing alt text for images inside a post. On a post that is already published, those are live immediately. The assistant reads the wording back and asks for a yes first, but there is no draft step in between.
Changing a post's URL handle leaves a redirect behind from the old URL, so links people already shared keep working.
It doesn't cost you AI tokens
Writing a post, suggesting keywords, generating metadata — inside the app that work spends your AI tokens. Through an assistant it doesn't: work you ask an assistant to do over MCP is not charged to your AI token balance.
Your balance is only for the AI features you run from the app itself. The assistant can still read the balance and tell you what's left, because that number is what the app charges against.
The approval screen
Signing in with ChatGPT or Claude ends on an approval screen. It names the client asking for access, names the store you're signed in to, and lists what the connection will and won't be able to do.
There is no permission to tick — the client asks for what it needs, the screen shows you what that is, and you either Approve or Deny.
Click Approve to finish. The connection then shows up on the MCP page with a Connected badge.
Manage or disconnect an assistant
Under AI assistants on the MCP page:
- See what's connected — ChatGPT and Claude.ai show a Connected badge and when they were connected.
- Reconnect — click Reconnect next to an assistant to run through the sign-in again.
- Disconnect — click the delete icon next to it and confirm. It loses access to your store straight away.
For tokens, see Manage your tokens above.
Disconnect any assistant and revoke any token you no longer use. Both stay valid until you remove them.
Troubleshooting
The MCP page shows a "Pro" badge and the endpoint is hidden
Your store is on the Free plan. The page stays visible so you can see what's there, but the Connect buttons are disabled and the endpoint is masked. Upgrade to Pro from the upgrade link at the top of the page, or from Subscription in the left menu, and everything unlocks.
My plan lapsed and now every tool refuses
That's the expected behaviour, not a broken connection. When a store drops off Pro, the tools refuse and tell the assistant to say so. Upgrade again and the same connection starts working on its own — there is nothing to reconnect and no token to replace.
Nothing happens when I click Connect
Your browser blocked the pop-up. Allow pop-ups for the Shopify admin and click Connect again.
You can also skip the button: copy the MCP endpoint from Setup guide and add the connector in your assistant yourself.
ChatGPT or Claude says it can't sign in, or the connector is broken
Two things to check:
- Your assistant's plan. Custom connectors need a paid plan on ChatGPT and on Claude. ChatGPT also needs Developer mode turned on.
- A stale connector. Remove the connector on the assistant's side and add it again — one created against an older setup keeps failing until it is recreated.
The assistant reads my posts but refuses to write one
My editor or terminal client is refused
Almost always the token. Work through these in order:
- Did you replace
YOUR_TOKEN? The config you copy from Setup guide is a template — the placeholder has to become the real token. - Is the token still there? Check the Access tokens card. A revoked token is gone from the list, and anything using it stops working immediately.
- Did you copy the whole thing? The token is shown once and is long. If it was truncated, revoke it and create a new one.
- Codex CLI only — the token comes from the
AVADA_BLOG_MCP_TOKENenvironment variable. Open a new terminal after adding theexportline to your shell profile.
Gemini CLI says connected but the assistant has no tools
Gemini CLI registers nothing from a server that isn't on its allow list, while still reporting it as connected. Keep the mcp.allowed block in your settings file — see Connect Gemini CLI.
I lost a token
There's no way to look one up — the app stores only a fingerprint, which is what you see in the Value column. Revoke it on the MCP page and create a new one, then update the client's config.
The assistant tells me a wrong number of posts
Ask it to count. Listing posts returns one page at a time, so a short page is not the end of the list. There is a separate counting tool that walks the whole blog — "how many posts do I have in total?" gets the exact number.
The assistant calls a tool that no longer exists
Reconnect it. A client keeps using the tool list it received when it connected, so it can keep calling something that has since been removed until it reconnects and picks up the current list.
I connected the wrong store
Each connection is bound to one store and there's no way to move it. Disconnect it on the MCP page of that store, then connect again from the store you meant.
If you run several stores, give each connector and each token a name that includes the store — the endpoint is identical everywhere, so nothing else in your assistant will tell you which store it points at.
Still stuck? Contact support from the app: open MCP → Setup guide → Other clients → Contact us, or use the chat bubble in the bottom-right corner of the app.