Skip to main content

Access and security

Beta

The MCP server is in beta. You can use it, but its tools, limits and behavior can still change. The changelog announces changes.

An AI assistant never gets more access than the user who connected it. The MCP server adds two limits of its own on top of the user's roles: the access the user grants at sign-in and a read-only cap for apps it doesn't recognize.

The assistant acts as the signed-in user​

Every tool call runs as the user who connected, with their Propeller roles. Propeller checks those roles on every request, as it does in the Backoffice. On top of that, the user chooses at sign-in what the assistant may do per module. The assistant gets the lower of the two:

User's role on Product DataGranted to the assistantEffective access
OwnerEditorEditor
EditorOwnerEditor
ViewerEditorViewer
OwnerNoneNone

The server only offers the tools the effective access covers, so an assistant with Viewer access never sees the tools that change data. A connection without a recorded grant gets read access.

One connection per person and app​

Each app a user connects, such as Claude, ChatGPT or Claude Code, has its own login, tenant and access. Connecting ChatGPT to another tenant leaves the Claude connection as it was, and disconnect ends only the connection it runs in. A connection never falls back to another connection's login.

Apps the server recognizes​

After sign-in the browser returns to the app that started it. Anyone can register an app and give it a familiar name, so the server checks the address the sign-in returns to:

  • Recognized: Claude (claude.ai and claude.com, the desktop app and Claude Code), ChatGPT (chatgpt.com), Microsoft Copilot Studio and Microsoft 365 Copilot, VS Code, Cursor and any app on the user's own device.
  • Any other address: the sign-in page warns the user and the app gets read access at most, whatever the user selects.

Sign-in and sessions​

  • The sign-in follows the MCP authorization specification: OAuth 2.1 with the authorization code flow and PKCE. Access tokens are issued for this server only.
  • The user signs in with their Backoffice account, including multi-factor authentication when their account uses it. The password goes to the Propeller sign-in only. The MCP server never receives it.
  • The server stores each connection's login encrypted.
  • A connection stays signed in for up to 90 days and ends after 14 days without use. A password change or a change to the authentication method ends it sooner. The assistant then asks the user to sign in again.

Audit trail​

Each sign-in is recorded with the user's email address, the tenant, the app, the access granted and whether the app is a recognized one. Each tool call is recorded with the user, the tenant, the app, the tool, the outcome and the duration. Arguments are recorded as a summary. Tokens, passwords and keys are never recorded.

Limits​

LimitValue
Tool calls60 per user per minute
Requests to the platform600 per user per minute. A bulk change that can't fit is refused before it changes anything.
Time per tool call50 seconds. Longer work stops and returns a partial result that says so.
Items per bulk change500 products and clusters, or 500 categories
Categories per create or delete call50, and up to 200 in one deleted branch
Categories per link_products call10 to add to and 10 to remove from
Items per reorder50

When a limit is reached, the tool says which one and when to try again.

Recommendations​

  • Grant only what the job needs. Read only is enough to explore a catalog.
  • Try changes on a test tenant first when you have one.
  • Read the previews and the confirmation prompts. Don't set your assistant to approve tools that change data automatically.
  • Disconnect a connection you no longer use.
  • The assistant sees what the tools return, such as category names and texts, product names and SKUs. Use it in line with your organization's policy for AI tools.

See also​