How AI Agents and the MCP Protocol Are Changing CMS and Website Builder Architecture
Anthropic
Webflow
uCoz
Sanity
Contentful
Framer
In 2026, agentic features are being massively integrated into CMS platforms and website builders, but not all of them are truly agents. We examine the architectural differences between generators, assistants, and agents, the role of the MCP protocol, as well as the risks and maturity levels of agentic platforms.
Habr publishes an article analyzing the integration of AI agents into CMS and website builders. The author notes that simply adding a "Ask AI" button does not make a platform agentic—often it is just a renamed text generation feature. A classification is introduced: generator (single request), assistant (suggests, human confirms), and agent (autonomously plans and executes multi-step tasks). Special attention is given to Anthropic's MCP protocol, which standardizes interaction between AI applications and data sources, but is not itself an agent. Architectural features of implementing agentic functions on different platforms are examined: headless CMSs (Sanity, Contentful) adapted faster; visual builders (Webflow, Framer) had to develop branching and action logs; enterprise solutions (Adobe Experience Manager) create separate agents for roles; and domestic platforms (uCoz) use crude but reliable control mechanisms (backup, HTML validation, rollback). Risks are also described: excessive permissions, partial execution of mass edits, working on the wrong project, prompt injection (up to data leaks via GitHub issue), and MCP server vulnerabilities (example of postmark-mcp sending copies of emails to a third-party address). An agentic maturity scale from 0 to 5 is proposed: most platforms in mid-2026 sit between levels 3 (limited tool set) and 4 (multi-step tasks with review).
Source: Habr — хаб ИИ —
original