Talos3D
AI-readable 3D design platform
Talos3D is building a different kind of 3D stack: one where geometry is not just pixels and meshes, but an authored model with structure, intent, and commands that both people and AI can understand. Recent work adds drawing views, dimensioning, guide-line inference, named views, multi-format paper export, and a persistent assistant lane that still operates through the same public model API as everything else.
cargo run --features model-api
# same command surface for UI and AI
view.front
view.apply_paper_preset
place_dimension_line
place_guide_line
export_drawingWhy it matters
When a model collapses into mesh-only state, people lose context and AI has to guess. Talos3D keeps authored entities, relations, views, dimensions, and commands legible so the model can stay understandable as it moves from sketching to editing to communication.
Rooms, storeys, openings, features, guides, and relationships do not need to be reverse-engineered from triangles.
The in-app assistant and external agents work through the same MCP and command surfaces instead of hidden editor hooks.
Recent drafting work turns the live model into orthographic views, dimensioned exports, and paper-ready PNG, PDF, and SVG output.
Key differentiator
Talos3D treats meshes, highlights, previews, and caches as derived artifacts. The real source of truth is the authored model: entities, definitions, parameters, semantic assemblies, typed relations, and evaluated facts.
Recent innovation
New orthographic drawing views, dimension offsets, guide-line inference, named views, and `export_drawing` turn modeling state directly into communicable paper-style output without a disconnected second workflow.
AI without special privileges
Model inspection, screenshots, view control, lighting, dimensions, guides, assemblies, and edits are exposed through a structured MCP interface backed by the same command and history pipeline the UI uses.
Broader than one vertical
The core already hosts modeling, architecture, terrain, content libraries, and assistant workflows. The capability model is designed so future domains can add their own tools, vocabularies, rules, and catalogs on the same footing.
Latest thinking
The current direction is clear: keep authored intent primary, expose the model publicly, let AI and automation operate on the same semantics as humans, and make views, drawing output, and domain workflows first-class instead of bolt-on extras.
Built with
Talos3D is grounded in a stack that is unusually well-suited to a long-lived 3D platform: Rust for correctness and performance, Bevy for ECS-driven runtime composition, and a browser-hosted direction that aligns with WebAssembly (Wasm) rather than locking the project to a desktop-only future.
Strong data modeling, reproducible command pipelines, and a credible foundation for serious geometry, import/export, and AI integration work.
An ECS architecture that maps cleanly to plugin composition, capability boundaries, authored entities, and multiple future workbenches.
Bundled catalogs, abstract storage boundaries, and managed assistant relays all point toward browser-hosted deployments that keep the same core model contract.
Open platform
If you care about AI-readable models, extensible domain workflows, or a cleaner bridge between 3D authoring and communication, Talos3D is worth watching now.