Skip to content

Security

AI clients can perform real actions in a QUIQQER installation. Treat every AI connection as an integration with production access.

Account Separation

Create dedicated QUIQQER accounts for unattended AI access. Interactive OAuth connections act as the user who approves the consent request, so that user should also have only the permissions required for the MCP workflow.

Do not use:

  • personal administrator accounts
  • shared team accounts
  • root or superuser accounts for normal editing workflows
  • one token for multiple unrelated AI clients

Use one account and one token per unattended workflow where possible. For interactive OAuth, review the MCP functions on the consent page before granting access. These practices make access review and revocation predictable.

Least Privilege

Grant only the permissions required by the workflow.

Examples:

WorkflowTypical permissions
Read contentCore MCP access and read access to the target project/sites.
Edit contentCore MCP access and edit access to the target project/sites.
Manage mediaCore MCP access and media permissions for the target project.
Manage bricksBricks MCP access and normal Bricks permissions.
Clear cacheCore MCP access and quiqqer.core.mcp.clearCache.

Do not grant cache clearing, delete permissions, package management, or broad administrator access unless the workflow explicitly requires it.

Token Handling

Treat API tokens like passwords:

  • Never paste tokens into prompts, chats, issue comments, pull requests, or screenshots.
  • Store tokens in a local secret store or environment variable.
  • Use different tokens for local machines, servers, and automation jobs.
  • Rotate tokens regularly.
  • Revoke tokens when access is no longer needed.
  • Revoke tokens immediately if they may have been exposed.

OAuth-capable clients manage their access and refresh tokens automatically. Do not copy those tokens into prompts, logs, or shared configuration files. Revoke the OAuth client or its tokens when the connection is no longer trusted.

Review Model

For write workflows, prefer a review step:

  • The AI reads the current state first.
  • The AI proposes the intended change.
  • A human confirms sensitive changes.
  • The AI applies the change.
  • The AI reads the result back.

Use this especially for public pages, navigation, media replacement, deleting content, cache clearing, and changes across multiple languages.

Released under GPL-3.0-or-later.