// the find
sparklabx/drawio-ai-kit
Teach your AI to draw correct, beautiful draw.io diagrams — declarative layout engine, ground-truth stencils, structural validator, vision self-check. AWS · Azure · GCP · Databricks · BPMN. Zero dependencies.
A zero-dependency Node CLI (`drawio-ai`) plus five agent skills that let an AI build draw.io diagrams for AWS, Azure, GCP, Databricks and BPMN. The agent looks up real stencil names and colors from a catalog instead of guessing them, lays out nested groups from a declared tree instead of coordinates, and runs a static validator before rendering. It is aimed at engineers who already have an agent drawing architecture diagrams and keep getting invented icon names or overlapping boxes.
- The AWS catalog is generated from a 10,446-shape index (jgraph/drawio-mcp, Apache-2.0), and `drawio-ai search` returns the exact stencil name and style string. That fixes the fabricated-stencil problem at the lookup step rather than asking the model to try harder.
- The layout engine takes a topology type and a nested group tree, sizes frames to their children, and routes edges. The examples contain no hardcoded coordinates, which is the right abstraction for a model that is poor at pixel arithmetic.
- The validator checks what can be checked statically: unknown stencils, dangling edges, recolored AWS icons, broken AWS group nesting, and children spilling out of their frames. Running these before the slower render-and-look step is the correct order.
- No postinstall hooks, no runtime dependencies, and no telemetry. For a CLI an agent shells out to, that is the right default and it is easy to audit.
- It installs from a GitHub ref, not the npm registry, so there is no semver resolution and pinning is manual (`#v1.0.1`). That is workable for an early project but awkward for reproducible installs in CI.
- "Zero dependencies" holds only for the Node side. Rendering needs the draw.io desktop CLI, large graphs need Graphviz, and the `logo` command and catalog regeneration need Python 3.11. The pack builder also needs macOS `qlmanage`, so contributors on Windows or Linux cannot rebuild the icon packs as documented.
- The aesthetic audit (font, palette, fan-out, icon size) is a set of heuristics. The vision check is done by the same model that drew the diagram. The README does not show how often that loop converges or how often the audit flags a diagram that is actually fine.
- Licensing is split. The code is MIT, but the bundled AWS Architecture Icons and third-party logos are not, and the base64 bundle path carries redistribution risk. Anyone redistributing an icon bundle needs to read NOTICE first.