Limits and behavior
- Every connection can read. A connection can make changes only if Make changes on your behalf was checked when it was approved.
- Changes are attributed to the person who approved the connection.
- A connection lasts as long as you chose when you approved it: 1, 7, 30 or 90 days, or until you disconnect it. A timed connection ends that long after approval, however often it’s used; after that the assistant has to connect again, and you approve it again. A connection kept until you disconnect it doesn’t expire.
- If the person who approved a connection loses access to the store (they are removed from it, or their Depict account is deactivated), the connection stops working at its next request. If they still have no access a day later, the connection ends and leaves Settings → AI assistants. Giving them access again doesn’t bring it back: they connect the assistant again.
Requirements
- A Depict account with access to the store. You sign in with it when you approve the connection.
- An assistant that supports remote MCP servers with OAuth sign-in. The steps below cover Claude, Claude Code, ChatGPT and Codex. Other assistants that work the same way can connect too.
Your endpoint URL
Every store has its own endpoint URL:- Your MCP endpoint shows the URL with a Copy button, and the full Claude Code and Codex commands, each with its own Copy button.
- Connected assistants lists the assistants connected to the store.
store_id part ties the connection to your
store; without it, the assistant can’t connect.
Connecting an assistant
Whichever assistant you use, connecting works the same way: you give it the endpoint URL, it opens the approval screen in Depict, you sign in and approve it, and it is connected. You don’t create keys or enter a client ID or secret.Claude
Claude on the web and the Claude desktop app share connectors: a connector added in one is available in the other. Custom connectors work on every Claude plan, but the Free plan allows only one. Who adds the connector depends on your plan:- Free, Pro and Max: you add it yourself.
- Team: an owner or primary owner adds it for the organization.
- Enterprise: an owner, a primary owner, or someone whose role includes Manage access to Libraries adds it for the organization.
1
Add a custom connector
In Claude, go to Customize → Connectors, click + Add, then Add
custom connector. On Team or Enterprise, the person who adds it for the
organization goes to Organization settings → Connectors instead and
clicks Add, then Custom, then Web.
2
Enter the endpoint
Give the connector a name, for example
Depict, paste your endpoint URL
and click Continue. Keep the OAuth sign-in Claude detects and leave the
OAuth client ID and secret empty. If the dialog asks how Claude should
identify itself, choose Register automatically. Then click Add.3
Connect
Click Connect on the connector in Customize → Connectors. On Team
or Enterprise, it’s listed there for everyone in the organization, usually
marked Custom. Your browser opens Depict; see
The approval screen.
4
Turn it on in a chat
In a chat, open the + menu, then Connectors, and make sure the
Depict connector is on.
Claude Code
1
Add the server
Copy the command from Claude Code on the AI assistants page and run it
in your terminal. It looks like this:Keep the quotes: without them, some shells treat the
? in the URL as a
wildcard.2
Sign in
Start Claude Code, run
/mcp, select depict and choose to
authenticate. Or run claude mcp login depict in your terminal. Your
browser opens Depict; see The approval screen.ChatGPT
ChatGPT connects to custom MCP servers on chatgpt.com in a browser, not in the mobile apps. You don’t need to turn on Developer mode. Who can add Depict depends on your plan:- Business: only a workspace admin or owner.
- Enterprise and Edu: admins and owners, and members an admin has given the Create plugins with MCPs permission.
- Pro: you add it yourself, but ChatGPT uses only the tools that read. It can answer questions about your store, not make changes.
1
Add a custom MCP server
In ChatGPT, open Plugins, click + and choose Add custom MCP
server. Give it a name, for example
Depict, and paste your endpoint
URL under Connection. Choose OAuth as the authentication and leave
the client ID and secret empty: ChatGPT registers itself with Depict.2
Create it and connect
Read the risk warning, check I understand and want to continue and
click Create as a plugin. When ChatGPT starts the sign-in, your browser
opens Depict; see The approval screen. After you
allow it, you are sent back to chatgpt.com, which lists the Depict tools.
3
Publish it for the workspace
On Business, Enterprise or Edu, an admin or owner publishes Depict from
the workspace’s admin console, under Plugins. People in the workspace
then find it under Settings → Plugins.
4
Use it in a chat
Start a new chat, type
@ and pick Depict. The pick applies to that
message, so type @ again when a follow-up needs new data from your store.
Naming it in your request helps, for example “Use Depict to list my
collections”.Codex
The Codex CLI, the Codex IDE extension and the Codex app share one configuration, so adding Depict once makes it available in all of them.1
Add the server and sign in
Copy the command from Codex on the AI assistants page and run it in
your terminal. It adds the server, then signs in:Keep the quotes, as with Claude Code. Your browser opens Depict; see
The approval screen. To sign in again later, run
codex mcp login depict on its own. In a terminal without a browser, add
--no-browser to the login command and follow its instructions.2
Allow
Approve the connection. The browser returns to Codex, which stores the
sign-in.
codex mcp list, or /mcp inside Codex, shows the Depict server
and its tools.The approval screen
When an assistant connects, your browser opens a Depict approval page. If you are not signed in, sign in first. The request waits for 10 minutes. The page reads [name] wants to connect to [store], followed by where approving sends you. Read both before you allow it. The name is what the assistant calls itself. Any program can register under any name, so the name alone doesn’t tell you who is asking. Where approving sends you is the part you can check. When you click Allow, your browser is sent back to the assistant with a one-time code that completes the connection. The page says where that is:
For a website or an app link, the page adds Only allow if you started this
from [name] yourself. If you didn’t just start connecting an assistant,
click Cancel: someone may be trying to get access to your store through a
link they sent you.
Choose what it may do:
- Read your store’s Depict data (collections, products, analytics) is required and always checked.
- Make changes on your behalf (publish curation changes, edit rules) is unchecked. Check it only if you want the assistant to make changes. It appears only when the assistant asked for changes.
What an assistant can do
A connected assistant gets these tools:- overview: a one-page explanation of how Depict models a headless store. The assistant reads it before its first search.
- search: finds the Depict operation that answers a question.
- execute: reads your store’s data. It never makes changes.
- execute_write: also makes changes. Only a connection you allowed to make changes gets this tool.
- “Which collections have the most hidden products?”
- “List every pinned product in Norway, collection by collection.”
- “Compare last week’s revenue per collection with the week before.”
- “Why doesn’t search show this product?”
- “Pin these three products at the top of the SUMMER SALE collection.”
- “Set the pins of BESTSELLERS in Finland to this list, in this order.”
- “Pin these products in the search start view, in this order.”
- “Hide these 40 products in the OUTLET collection.”
- “Move the SUMMER SALE collection under OUTLET.”
- “Translate the filter word Colour into Swedish.”
- “Show the Size filter before the Colour filter.”
- “Add jumper as a synonym for sweater.”
- “Turn off the query rule for rain jacket.”
- “Hide these three products from search in every market.”
- “Reset these 31 products to From feed.”
- “Add linen shirts to the Swedish search suggestions.”
- “Show what You may also like would show for this product with only navy products.”
- “Move Finland from the Nordics market group to Europe.”
- An answer over 1 MB is refused with a message to narrow the request. Ask for fewer rows, a shorter period, or leave out extra metrics.
- Ask for only a few rows of product metrics at a time: at most 40 products over 30 days, 15 over 90 days or 5 over a year.
- Metrics reads can take 10 to 30 seconds. Ask for a shorter period if one fails.
- Comparing two periods for one market or market group covers at most 10 collections. The assistant tells you when that happens, and the totals add up only those collections. To choose which collections, name up to 10 of them.
- A bulk hide, unhide or hand-picked change takes up to 500 products at a time.
- If you name something Depict doesn’t recognise, such as a mistyped product id or a deleted market group, you still get an answer, and the assistant tells you what wasn’t recognised. A product Depict has no record of shows as visible, and for a market group that doesn’t exist or has no markets, the numbers cover the whole store, not that group.

