Diagrams are one of the last places where "just describe what you want" still fails, because the output isn't prose — it's strict draw.io XML where one malformed tag breaks the whole canvas. This project's bet is that frontier models are now reliable enough at long-form structured generation to write that XML directly, so the chat box becomes the diagramming tool and the canvas just renders what the model emits.
What Sets It Apart
- The diagram is the model's output format, not a side effect. Commands map to generated or modified draw.io XML rendered through react-drawio, so what you edit by hand and what the AI edits are the same artifact — no lossy round-trip between a chat representation and the canvas.
- Ingest, then reconstruct. Drop in an image, PDF, or text file and the model extracts its structure to rebuild it as a real, editable diagram rather than a static import.
- Bring-your-own-everything. Vercel AI SDK fronts a long provider list (OpenAI, Anthropic, Google, Bedrock, DeepSeek, Ollama, OpenRouter, Doubao) with custom endpoints, an admin panel for model/quota management, and an MCP server so Claude Desktop, Cursor, or VS Code can drive it as an agent.
Who It's For
Great fit if you live in draw.io and want a conversational shortcut for first drafts and bulk edits — especially cloud architecture diagrams — while keeping a fully editable canvas and version history to fix what the model gets wrong. Look elsewhere if you need deterministic, pixel-perfect layouts or want to avoid depending on a strong (often paid) model: reliability rests on the LLM's ability to emit valid XML under formatting constraints, so weaker models produce broken or drifting diagrams.
