lightningpixel/modly

Modly runs image-to-3D mesh generation entirely on your local GPU, no cloud API key required. You feed it a photo; it produces a 3D model. The architecture is an Electron desktop shell over a FastAPI Python backend, and the whole stack runs on Windows, Linux, and Apple Silicon macOS.

The model support comes through an extension system rather than being baked in. Each extension is a GitHub repository with a manifest.json and runtime entry files. The five official extensions cover Hunyuan3D 2 Mini (plus turbo and fast variants), TripoSG, and Trellis2 in GGUF format. Installing one means pasting a GitHub URL into the Models page - the app handles the rest. That separation matters: the core application doesn't bloat with every model weight, and third parties can ship their own extensions against the same contract.

Workflows are node graphs. The README's suggested starting point - Image → Generate Mesh → Add to Scene - is deliberately minimal, and the app validates graph wiring before execution rather than silently dropping your current mesh view when something is misconfigured. Inline and toast warnings surface instead. Imported meshes can also be smoothed and decimated in-app, with results written back into the workspace.

The part most relevant to anyone building automated tooling is the CLI. tools/modly-cli/agent.py is a stdlib-only Python script that talks to a running Modly desktop instance over HTTP. The canonical surface is small: health, model, workflow-run, capability, process-run, and the convenience generate command that posts to /workflow-runs/from-image, polls the run, and exports the final mesh as a GLB. The JSON on stdout includes recovery metadata - status and cancel commands - so a script can handle failures without manual intervention. The README is explicit about what is and isn't canonical: legacy wraps old job endpoints, dev serve-api only starts the FastAPI backend without proving the Electron bridge is ready, and the experimental ComfyUI helpers are outside the main agent contract. That kind of intentional surface taxonomy is rare in a project this early.

The macOS constraint is Apple Silicon only - no Intel support. The live RAM indicator in the top bar suggests the developers are aware that local model inference is memory-hungry and users will want to watch it. The MIT license includes an explicit attribution requirement for forks, which the README states plainly.

For agent builders specifically, the combination of a local inference backend, a well-defined CLI contract with machine-readable JSON output, and a node-graph workflow model makes Modly a plausible component in a larger pipeline - not just a standalone desktop toy.