Eraser opens its diagram format under MIT: models can now say where each box goes

Eraser has open-sourced its diagram format under the MIT licence. This is not a new drawing tool. It is a format, and its point is that an AI describing a diagram should also be able to say where the boxes go.

The problem

Until now AI diagram tools, Eraser's own included, let the model say only what connects to what and handed layout to an engine. In Eraser's words: "Mermaid is a relationship language. It has no way to say where anything goes, layout is delegated to an engine by design." The same holds for D2 and Graphviz.

That was a choice, not a flaw, because models were poor at translating design intent into coordinates. The output was overlapping nodes and spaghetti lines. According to Eraser this changed in late 2025 and early 2026: frontier models became genuinely fluent at coordinate-based layout. The new capability then had nowhere to land, because no format could carry it.

Why not an existing format

Eraser rules out the alternatives one by one. SVG and HTML are too low-level, with no semantic notion of nodes and edges. Draw.io XML is too verbose and token-expensive. Mermaid, D2 and Graphviz cannot express coordinates. The opened format fills the gap between them: semantic nodes and edges, plus explicit coordinates, plus strong defaults, in a syntax a model can emit cheaply.

Why open it

The company's reasoning: if human comprehension is the bottleneck in human-agent collaboration, the fix has to have maximum reach. Everywhere an agent needs to show a human its work, or help someone grasp a complex system.

The licence is MIT, so commercial use is free, though the repository is kept closed to outside contribution. The announcement gives no adoption or performance figures.

Sources: Eraser blog post, github.com/eraserlabs/eraser-diagrams.