How AI Agents and MCP Protocol Change CMS Architecture and Website Builders

The article examines the impact of AI agents and the Model Context Protocol on CMS and website builder architecture. It distinguishes between generators (single‑prompt output), assistants (context‑aware suggestions), and true agents (autonomous planning and execution). MCP, introduced by Anthropic in late 2024, standardizes how AI applications connect to tools and data, reducing the N×M integration problem. The author proposes a six‑level autonomy scale: from Level 0 (no AI) to Level 5 (closed‑loop, no human confirmation). Currently most platforms are at Level 3–4. Architectural differences explain adoption order: headless CMS (Sanity, Contentful) integrated agents first, followed by visual builders (Webflow, Framer), then enterprise platforms (Adobe Experience Manager), and finally niche builders like uCoz. Key risks include excessive permissions, mass partial updates, environment confusion, prompt injection via untrusted content (e.g., hidden commands in GitHub issues), leaked authentication tokens (the postmark-mcp npm package case), and insecure MCP server code (CVE-2025-6514). Platforms must treat CMS content as untrusted data and implement auditing, versioning, and explicit confirmation for agent actions.