Modly

Why it's ranked here

Few local image-to-mesh tools are this complete. Modly ships installers for Windows, Linux and Apple Silicon Macs, five official model extensions including Hunyuan3D 2 Mini, TripoSG and Trellis2, and three automation surfaces for agents. The engineering shows care in the places that usually break: workspace folder names are checked so output cannot escape the workspace, cancelling a job kills the model process so the GPU stops at once, and a written design decision covers memory limits on 16 GB Macs. MIT licensed, with Python and Node test suites wired into the test command.

What's good

Model isolation is the strongest choice. Each extension gets its own environment and talks to the app over plain line-by-line JSON, so one model's dependencies cannot break another, and unloading means ending the process, which the Mac design notes call the only reliable way to free GPU memory. Auto-repair of a missing package is deliberately narrow: it maps only known module names rather than guessing. Live progress bars from models are forwarded character by character so the log pane stays current during long phases. The command-line helper returns machine-readable JSON with run status and cancel hints for recovery.

Tradeoffs

The local web service allows requests from any browser origin and serves workspace files by path, so anything running in a local browser can reach it while Modly is open. Installing an extension means running that repository's own model code on your machine. Mac builds are Apple Silicon only, with Intel and Rosetta out of scope, and the packaging scripts switch off automatic code-signing discovery. The MCP server still drives the older job endpoints the README labels legacy. NVIDIA Jetson is not officially targeted; the community guide needs a hand-swapped PyTorch, older NumPy and a patched background remover. The license adds a request to credit Modly visibly in any fork.

How to use it well

Pick it if you want photo-to-mesh generation on your own GPU and are happy to install models from GitHub repositories you trust. Start with the simple image, generate mesh, add to scene workflow and check the logs panel when something fails. For pipelines, use the command-line generate command, which starts a run, polls it and exports the mesh in one call. Agent users can point an MCP client at the bundled server, but the app must already be running. The backend also runs headless over HTTP. It does not cover text-to-3D in the fetched code, and textures need extra native builds on Jetson.

Observed

License

MIT, with an added request to credit the original project visibly in forks

Language

TypeScript Electron and React desktop shell with a Python FastAPI backend

Packaging

Windows NSIS installer, Apple Silicon macOS dmg, Linux AppImage; launch scripts for running from a clone

Platform support

Windows, Linux and Apple Silicon macOS; Intel macOS out of scope; Jetson unofficial via a community guide

Interfaces

Desktop GUI, local HTTP API, stdlib-only Python CLI, MCP server over stdio

Extension model

Model extensions are GitHub repositories with a manifest, run in their own venv as a subprocess speaking newline-delimited JSON

Tests

npm test runs both a Python test suite and Node test files

Local API CORS

Backend allows all origins, methods and headers