Terrastruct Open-Sources TALA, Its Diagram Autolayout Engine, Under MPL-2.0

Terrastruct has released TALA, the autolayout algorithm behind its D2 diagramming language, as open source under MPL-2.0. Unlike the DAG-oriented engines D2 already ships (Dagre and ELK), TALA is built for software architecture diagrams and favors orthogonal layouts that resemble whiteboard sketches rather than graphs that grow in a single direction. It draws on published graph-drawing research plus original techniques, and optimizes for several aesthetic goals at once—symmetry, median node distance, flow, and clustering of similar nodes.

A notable feature is support for fixed and partial node positioning: callers can lock specific coordinates while TALA places the rest and handles edge routing. Terrastruct pitches this at agentic workflows, arguing that language models can reason about 2D placement but struggle with routing, so a human or model can pin key nodes and let the engine finish the layout. Several of the demo compositions were AI-generated.

The tradeoffs are stated plainly. TALA uses randomized seeding (three by default, keeping the best-scored result), so adding a single node can reshuffle an entire diagram in ways Dagre and ELK would not. It handles long DAG-style flows less gracefully and scales nonlinearly in runtime on larger diagrams. TALA ships in D2 v0.9.0 and can be enabled with —layout=tala or tried in the fully client-side web playground.