An MCP remote control for Draw.io diagrams

What it does

The Draw.io MCP server exposes the diagram editor to any Model Context Protocol client—Claude, Zed, Codex, and others—through tools that create, read, update, and delete diagram elements. It either hosts a built-in Draw.io editor locally or attaches to an existing browser tab via an extension, letting an AI assistant build architectural diagrams, flowcharts, and visual documentation from conversation.

The interesting bit

Vendor stencils for AWS, GCP, Azure, Cisco19, and CiscoSafe are auto-discovered at runtime from Draw.io’s own sidebar, so agents can place icons like mxgraph.gcp2.cloud_run without a hand-curated catalog. Live operations are serialized per document with a FIFO queue, which keeps multiple agents from colliding when they work on different files simultaneously.

Key highlights

• Built-in editor runs without a browser extension; an extension is only needed to connect to an existing browser tab.
• Supports multi-document and multi-page targeting, plus page management tools such as create-page, copy-page, and rename-page.
• Imports and exports XML, SVG, and PNG (all with embedded XML), and can embed or expand Mermaid diagrams.
• Optional TLS mode with auto-generated self-signed certificates via a per-user local CA.
• Vendor icon stencils are discovered at runtime rather than hard-coded.

Caveats

• Integration with the draw.io desktop Electron app is experimental and currently blocked by an upstream Content Security Policy issue that rejects the WebSocket connection.
• Firefox support requires running the server in TLS mode.

Verdict

Worth a look if you want AI assistants to generate or maintain living architecture diagrams alongside your code. Skip it if you need the draw.io desktop app, since an upstream CSP issue keeps that integration experimental and non-functional.