Core dna MCP via Claude Code (CLI)

Add an MCP server
Run this command from inside the site's working directory in your terminal.

claude mcp add --transport http your-production-website https://www.yoursite.com/cdna-mcp -s project
This command must be run in your terminal inside the site directory - not inside a Claude prompt.
It writes to a local .mcp.json file which stores no credentials.
Using -s project scopes the MCP to that directory only, so it loads only when you work on that customer's site. This is the recommended approach

1. Hostname
Replace www.yoursite.com with any PROD or DEV hostname for the target site.

2. Server name
Use a descriptive name like your-production-website or client-dev. If you add both PROD and DEV MCPs, the names must be clearly distinguishable in prompts.

3. Verify and authenticate
Start Claude > run /mcp, select your new server, then choose:
Authenticate
An OAuth browser flow will open - log in with your credentials.
Or if already logged in click "Allow" to authorize Claude Code to access Core dna . 

Remove an MCP server
Use the scope flag that matches how the server was originally installed.

claude mcp remove your-production-site -s user
(Removes from your global user config)

claude mcp remove your-production-site  -s local
(Removes from local machine config)

claude mcp remove your-production-site  -s project
(Removes from the project's .mcp.json)

⚠️ Warning: AI + MCP Access to Your CMS/eCommerce Platform

Connecting AI agents to your CMS, eCommerce platform, or production systems through MCP (Model Context Protocol) introduces powerful automation capabilities - but also significant operational risk.

AI systems execute actions based on prompts and available permissions. A poorly written, ambiguous, or unintended prompt can trigger destructive actions at scale, including but not limited to:

• Deleting products, pages, collections, or customer data
• Publishing incorrect or unapproved content
• Modifying pricing, inventory, or SEO settings
• Overwriting templates, themes, or configurations
• Sending unintended customer communications
• Triggering irreversible bulk operations

Unlike traditional software workflows, AI agents interpret natural language probabilistically. Even well-intentioned prompts can produce unexpected outcomes if safeguards are not in place.

To reduce risk, we strongly recommend:

• Restricting AI permissions to the minimum required access
• Using staging or sandbox environments whenever possible
• Requiring human approval for destructive or bulk actions
• Implementing backups, versioning, and audit logs
• Testing prompts carefully before production use
• Monitoring all automated AI activity

By enabling MCP-connected AI access, clients acknowledge that AI-generated actions may produce unintended changes or operational disruption if improperly configured or prompted.

Treat AI agents with the same caution as granting administrator access to a human operator - because, at scale, mistakes can cause significant damage within seconds.