Agentic Mermaid inverts a basic assumption most diagram tools make: that humans are the primary input layer. Instead, it treats AI agents as first-class creators, letting them programmatically generate and edit Mermaid diagrams while still producing polished exports across 15 diagram families.
The interesting move isn’t the programmatic angle alone. It’s the verification loop, which runs structural and lint checks before export. That suggests the creators are thinking genuinely about how agents and humans collaborate: an agent drafts, a human reviews the readable Mermaid source, the agent revises. The workflow is built for iteration across multiple passes.
The MCP interface, one source, three channels (library, CLI, MCP), is also smart. MCP is becoming the standard way agents reach external tools, so making that a first-class distribution channel is the right move for 2026.
The repo backs that up: github.com/adewale/agentic-mermaid is MIT-licensed with a CHANGELOG, ships to npm as agentic-mermaid, and is forked from lukilabs/beautiful-mermaid rather than built from scratch. The real contribution sits on top of that renderer: typed mutation helpers, the verification loop, the MCP server. The GitHub handle isn’t tied to a named person on the site, a minor gap next to an open, actively maintained project.
The inversion and the verification loop are still the real ideas here. The repo link was just sitting in the footer the whole time.