Licensing clarification for commercial use of OmniParser & OmniTools and all its components: YOLOv9 / Ultralytics dependencies, PR #37, and host-machine deployment

I am looking to use OmniParser and OmniTool for a closed-source commercial project, but I want to get absolute clarity on the licensing before moving forward to avoid any legal roadblocks. Could you please help clarify a few specific concerns?

1. Host Machine Deployment vs. Virtual Machine

While OmniTool's documentation and examples often revolve around controlling a Windows 11 VM, can OmniParser and OmniTool components be run directly on a host machine (e.g., local development or production server) for our workflow, or are there hard dependencies tying OmniTool strictly to a VM environment?

2. Ultralytics Dependency vs. Main Branch

I noticed that ultralytics==8.3.70 (which is licensed under AGPL-3.0) is still listed in the requirements.txt file. However, looking at the main branch, it seems the repository has shifted towards using a YOLOv9-E detector (such as the weights from [Hugging Face PR #37](https://huggingface.co/microsoft/OmniParser-v2.0/tree/refs%2Fpr%2F37)).

• Is the remaining ultralytics entry in requirements.txt a leftover artifact from older versions?
• If someone completely avoids importing Ultralytics and sticks to the updated weights, does it completely decouple the project from AGPL restrictions?

3. Origin and Licensing of the YOLOv9 Implementation

The README mentions that icon_detect_v3 uses an MIT-licensed YOLOv9 implementation. Could you provide more details on this exact version?

• Where is it sourced from, and is it considered the "real" official YOLOv9?
• Do you use LibreYOLO, or is it based on this version: [MultimediaTechLab/YOLO Issue #231](https://github.com/MultimediaTechLab/YOLO/issues/231)? (I ask because that particular implementation has raised concerns in the community regarding matching official YOLOv9 benchmark scores on the COCO validation set).

4. Status of Hugging Face PR #37 Elements

In the main branch, elements and weights are referenced from [Hugging Face PR #37](https://huggingface.co/microsoft/OmniParser-v2.0/tree/refs%2Fpr%2F37). Are all files, models, and scripts under this PR/branch fully MIT-licensed?

Summary Request: Definitive Answer

To give our legal and engineering teams peace of mind: Can we safely use OmniParser and OmniTool for a closed-source commercial project right now without running into any viral licensing (AGPL) issues, or are there still components/weights that require caution?

Thank you so much for your time and guidance!

Start by reading README and requirements.txt, then inspect the main branch’s icon_detect_v3 references and the files linked from Hugging Face PR #37. Verify which dependencies, weights, and scripts are actually used and document their stated licenses and host-machine or VM assumptions. Done means a maintainer-approved clarification addresses each licensing question.