D2 Goes Non-Profit, Open-Sources TALA, Commits to AI-Only Development

Overview: D2's Strategic Pivot in One Week

D2, a Go-based text-to-diagram tool often positioned as a Mermaid competitor, announced two major changes in September 2026: its parent company Terrastruct ceased operations and donated the project to Hack Club (via The Hack Foundation, a 501(c)(3) non-profit) for fiscal sponsorship, making D2 a fully open-source, non-profit project; simultaneously, the previously paid core asset — the TALA (Terrastruct's AutoLayout Algorithm) orthogonal layout engine — was open-sourced under the same MPL-2.0 license as the main project, shipping with D2 v0.9.0.

From Open-Core to Fully Non-Profit

D2 previously followed an open-core model: the language and compiler were free, while revenue came from the closed-source D2 Studio IDE and the TALA layout algorithm. With Terrastruct shutting down, the founder chose to decouple the project's survival from any single company's commercial fate. Hack Club provides fiscal sponsorship; donations are tax-deductible in the U.S., funds are publicly traceable, and neither founder draws a salary from the fund. Development continues publicly on GitHub (github.com/d2lang/d2), and the founder notes that high-quality community maintainers could eventually be paid via donation-funded contracts.

Governance and Funding Structure

The sponsorship page (d2lang.com/sponsor/) confirms The Hack Foundation handles finances. The founder emphasizes that sponsorship does not influence roadmap, governance, or licensing. This mirrors prior transitions (e.g., projects moving to Linux Foundation or Apache) but is notable for its transparency and immediate full open-sourcing of the commercial differentiator.

AI-Driven Development Declaration

"This phase of D2 relied on pure hand-written code. That era is over. The next phase of D2 will be AI-driven — all my code contributions will use AI, and I welcome community AI-generated contributions, which I will review with AI. The sole exception is written content: all blog posts and external communications I will write myself, without AI."

The founder cites efficiency: in the past year he wrote zero lines of code by hand while still shipping features. He acknowledges controversy, explicitly referencing Neil Alexander's June 2026 post "Please stop flooding our projects with AI slop to furnish your CV" (neilalexander.dev/2026/06/30/flooding-contributions), which describes low-effort AI-generated PRs suspected of resume-padding. The founder accepts the risk and criticism, framing the move as a transparent experiment rather than a recommendation.

Three "Non-Goals": A Restrained Product Philosophy

No bloated all-in-one diagram tool. AI could easily turn D2 into a Swiss-army knife, but the founder prefers spending "compute budget" on a small, sharp core. Example: Venn diagrams are explicitly out of scope for the core; a future plugin/module system would host such extensions.

No built-in LLM interface. Current models can already convert requirements, code, or specs into .d2 files given a few examples; a dedicated chat UI inside D2 is unnecessary.

No server-dependent capabilities. As a non-profit, D2 will not run services: no MCP, no API, no user accounts, no server-side rendering, no real-time collaboration. D2 Studio will become a pure client-side, offline-capable application.

Roadmap: Three Focus Areas

Isometric renderer (previously blocked by engineering capacity).

Transpilers from other diagram languages (e.g., Mermaid, PlantUML) to D2.

Hybrid layout for the agent era: LLMs place high-level blocks; TALA handles aesthetic routing and detailed placement.

The founder argues that LLMs are good at rough 2D placement but weak at simultaneous constraints (no crossing edges, even spacing, global coherence). A text-based "source of truth" (like HTML for web) remains essential; D2 aims to be the standard substrate for "text-to-graphics" in the agent era.

TALA Deep Dive: Orthogonal Layout for Architecture Diagrams

What Is TALA?

TALA (Terrastruct's AutoLayout Algorithm) is a custom orthogonal layout engine developed over three years. Unlike Dagre and ELK (both layered/DAG-based), TALA produces layouts resembling hand-drawn whiteboard architecture diagrams. It optimizes multiple aesthetic metrics — symmetry, median node distance, flow direction, clustering of similar nodes — blending ideas from academic graph-drawing papers (cited in the code) with original techniques.

Usage: d2 --layout=tala architecture.d2 architecture.svg or try it instantly in the browser at play.d2lang.com (fully client-side, no server round-trip).

TALA vs Dagre vs ELK: Comparative Analysis

The official blog compared all three engines on seven real-world public D2 files from GitHub, using identical source and compiler, only switching the layout flag. Results are presented side-by-side with the admission that "for some diagrams you may prefer the non-TALA layout." Key dimensions:

Layout style: Dagre/ELK — layered, unidirectional flow; suits long-chain diagrams. TALA — whiteboard-like orthogonal; suits system-overview diagrams.

Stability: Dagre/ELK — local changes on node add/remove; overall shape stable. TALA — randomized (3 seeds, best score picked); same seed = same output, but adding one node can trigger global re-layout.

Coordinate control: Dagre/ELK — algorithm decides. TALA — supports partial/full manual locking of top / left coordinates; algorithm places remaining nodes and routes all edges.

Best-fit scenarios: Dagre/ELK — flow-oriented, long-chain, DAG-heavy diagrams. TALA — system-landscape diagrams needing symmetry; ideal for AI-generated block placement + algorithmic cleanup.

Performance: Dagre/ELK — near-linear scaling. TALA — non-linear slowdown on large graphs; benchmark repo: github.com/d2lang/d2-benchmarks.

Coordinate Locking: The Agent-Ready Feature

TALA's ability to accept fixed top / left coordinates for a subset of nodes is highlighted as critical for AI-agent workflows. The model proposes high-level block positions; TALA computes the rest and routes edges cleanly. Example from the blog: "The Printing Room" diagram locks four CMYK print stations in a row with equal spacing (mechanical alignment), while ten other nodes (feeders, cameras, controllers, dryers) are auto-placed.

Self-Declared Limitations

Randomness. Global re-layout on minor changes can be a drawback where visual stability matters.

Weak on pure DAGs. The founder admits preferring Dagre/ELK for long, single-direction flow diagrams.

Non-linear performance on large graphs. Benchmarks provided for independent verification.

Why This Matters for Developers

Immediate user benefit. Previously paid features (Studio enhancements, TALA) are now free; coordinate locking is especially valuable for teams automating architecture-diagram generation with AI.

Open-source governance case study. A concrete example of decoupling project continuity from a single vendor via fiscal sponsorship, with transparent finances and no IP transfer to a foundation.

Public AI-native development experiment. The founder's candid stance — embracing AI coding while drawing hard boundaries (no server, no bloat, no LLM wrapper) and pre-empting quality/trust concerns — offers a rare, observable template for other maintainers.

Next Steps

Upgrade to D2 v0.9.0 and test --layout=tala on existing diagrams. Monitor the GitHub repo (github.com/d2lang/d2) for the promised plugin system and isometric renderer. Donations accepted at hcb.hackclub.com/d2 with full public ledger.