What Happened
A developer named BV-Venky released Excalidraw Architect MCP, an open-source MCP (Model Context Protocol) server that generates properly laid out Excalidraw architecture diagrams from AI-powered IDEs like Cursor and Windsurf. The project, shared on Hacker News on March 7, 2026, tackles a specific and persistent problem: when AI coding assistants try to create Excalidraw diagrams, they hallucinate coordinates, causing boxes to overlap and arrows to cross into unreadable tangles.
The core of the tool is a layout engine that handles automatic positioning. You describe the architecture you want in natural language through your AI IDE, and the MCP server returns a clean Excalidraw file with proper spacing, aligned components, and readable arrow routing. No manual dragging required.
The server works with any IDE that supports the MCP protocol, which at this point covers most of the major AI coding tools. The project is available on GitHub and the developer notes the layout engine is already producing solid results, though it is still in early development.
Why It Matters
If you have ever asked Cursor or another AI IDE to generate an architecture diagram, you know the pain. Mermaid output is functional but ugly. ASCII art falls apart with anything beyond three boxes. And raw Excalidraw JSON is a disaster because LLMs simply cannot reason about 2D coordinate systems reliably.
This matters for anyone doing technical documentation, system design reviews, or onboarding materials. Architecture diagrams are one of those tasks that feels like it should be trivial for AI but consistently produces garbage. A dedicated layout engine that sits between the LLM and the diagram format is the right architectural approach. The LLM handles the semantic understanding of what components exist and how they connect. The layout engine handles the spatial math.
For teams already using MCP-compatible tools, this drops in without changing existing workflows. You keep working in your IDE, keep talking to your AI assistant, and now the diagram output is actually usable.
Our Take
This is a smart solution to a real problem, and it follows the pattern we keep seeing with MCP: specialized servers that handle the things LLMs are bad at while letting the model do what it is good at. LLMs understand system architecture. They cannot do spatial layout. Splitting those concerns is exactly right.
The practical value depends on how well the layout engine handles complex diagrams. Three microservices with a database? Easy. A full distributed system with 20 services, message queues, load balancers, and multiple data stores? That is where layout algorithms earn their keep. The early examples look promising, but real-world architecture diagrams get messy fast.
Worth watching if you do any amount of technical diagramming. The fact that it is open source and MCP-based means it will likely get contributions from the community and work across multiple IDEs without vendor lock-in. If you are currently copy-pasting Mermaid diagrams into Excalidraw and manually rearranging everything, this could save you 15-20 minutes per diagram.