Codex draws scientific figures live inside draw.io

What it does This is a Codex plugin with two MCP servers that launch the draw.io desktop app and manipulate its active graph model through the editor’s own API. The agent can add shapes, edges, and labels to the visible canvas with deliberate pacing, take screenshots to review its work, adjust styles and positions, and only save once the figure actually exists in the editor. A second server validates saved files and exports them to PNG, SVG, PDF, or JPG.

The interesting bit Most AI diagram tools generate an XML file and pray it renders correctly; this project uses Electron remote debugging to puppet the live draw.io canvas directly. That means the agent can see what it drew, iterate visually, and keep everything editable rather than shipping a static blob.

Key highlights

• Live canvas control via draw.io’s internal graph API, not OS mouse automation or pre-baked XML.
• Iterative workflow: the agent can screenshot, inspect cells, update styles, and fit the view as it builds.
• Exports to standard scientific figure formats (PNG, SVG, PDF, JPG) after validating the saved document.
• Binds only to localhost and refuses to attach to arbitrary browser pages; no telemetry or hosted backend.
• Ships as a complete Codex plugin with a bundled skill that teaches the agent pacing and visual review.

Caveats

• Windows is the only fully tested platform; macOS and Linux support is best-effort and may behave differently depending on draw.io’s Electron packaging.
• The live API is limited to draw.io primitives, so dense microscopy images, heatmaps, or complex plots are not yet handled well.
• Visual fidelity depends on how easily the reference can be decomposed into editable shapes and arrows.

Verdict Worth a look if you generate scientific figures with AI and want editable, inspectable output inside a real diagram editor. Skip it if you need photorealistic or raster-heavy illustrations, or if you are not already using Codex with plugin support.