# Cua perception third-party notices This document summarizes the inputs of the published `cua-perception` releases. It is not a substitute for the `NOTICE`, model ledger, source ledger, SBOM, and license texts shipped with an exact `cua-perception` artifact. If those authenticated records are missing or disagree with this summary, do not publish or redistribute the artifact. The default Cua Driver remains a separate MIT-licensed component and does not include the extension, model weights, or ONNX Runtime library. ## Before you install, redistribute, or host the extension This section summarizes known precautions. It is not legal advice. Consult your own counsel before redistributing the extension or offering it as part of a service. - **The Driver license does not cover the extension.** Cua Driver stays MIT licensed with or without the extension, because it talks to the extension worker as a separate process over a protocol. The extension artifact has its own licenses, and its OmniParser detector is AGPL-3.0-only. Cua does not relicense that detector and does not describe it as MIT. - **The extension is opt-in.** A default Driver installation never installs the extension or downloads model weights. If your organization does not accept AGPL-licensed components, do not install the extension. Driver keeps working and `parse_visual_regions` returns `not_installed`. - **Private use is unrestricted.** Running the extension on your own machines does not trigger the AGPL's source-distribution obligations. - **Redistribution carries the AGPL obligations.** If you copy the extension artifact to anyone else, include the AGPL-3.0 license text, the notices, the model and source ledgers, the SBOM, and the corresponding source, including the pinned source model and the conversion/export material listed below. - **Hosted services can trigger the network-use clause.** The shipped ONNX file is a converted, and therefore modified, version of the upstream model. If users interact with that model remotely through your service, AGPL-3.0 section 13 can require you to offer those users its corresponding source. This applies to any product that embeds Cua Driver with the extension enabled. - **The corresponding source has an upstream limit.** Cua provides the pinned upstream checkpoint, the conversion recipe, and the export script. The upstream model repository does not publish training data or training code, and Cua cannot supply them. - **The upstream origin is an Ultralytics YOLO model.** The OmniParser icon detector is a fine-tuned YOLO model, and its AGPL-3.0 terms come from Ultralytics. Ultralytics sells separate commercial licenses. Cua cannot grant any right beyond the AGPL-3.0 terms, so you need your own agreement with the rights holders if you need different terms. - **No warranty.** The extension and its model artifacts are provided without warranty, as their licenses state. Detection output can be wrong. Do not use it as the sole basis for irreversible actions. ## OmniParser icon detector - Origin: `microsoft/OmniParser-v2.0` model repository. - Pinned revision: `6600256cb0f1b07651e3bc86166196307bad7e2d`. - Source model: `icon_detect/model.pt`, SHA-256 `dab3d4351ad00b035db829909a4db98354d5a90f6990e4ac00222a9a95d4bf57`. - Reviewed converted ONNX artifact: SHA-256 `d8a876bf7f9fb73d7da9432904ade7fa78e092e9a91674e5a2806b45562a9ab2`. - Artifact license recorded by the release ledger: AGPL-3.0-only. On 2026-09-24 the Cua maintainers decided to distribute this artifact under AGPL-3.0-only, and the release ledger marks it `release-verified`. This is a maintainer decision, not a legal opinion. Cua Perception packaging refuses an AGPL model that lacks its bundled source input. A distributor must preserve the AGPL-3.0-only license and notices and provide the corresponding source and pinned conversion/export material required for this artifact. Network use of a covered modified version may also trigger the license's corresponding-source requirements. The artifact must not be presented as MIT merely because Driver invokes it through an extension protocol. The OmniParser source repository's license for repository content and the separate model repository's artifact terms are different records. Verify the exact model revision above. ## PP-OCRv5 The reviewed OCR inputs are official PaddlePaddle model repositories and are recorded as Apache-2.0: | Role | Repository revision | Reviewed ONNX SHA-256 | | --- | --- | --- | | Mobile detector | `PaddlePaddle/PP-OCRv5_mobile_det_onnx` at `e6f4fa85f00e168c862bc462aebca69eef9b3d3d` | `a431985659dc921974177a95adcfbb90fd9e51989a5e04d70d0b75f597b6e61d` | | English recognizer | `PaddlePaddle/en_PP-OCRv5_mobile_rec_onnx` at `3fafbc3b5dcf93dd72add9f48368be8a3a2cd33b` | `b5f833dfc5d0eb71da397b4efa06ebeee9b431b690a47d6af40d77d8eabc557f` | The recognizer dictionary comes from the pinned recognizer repository's `inference.yml`, SHA-256 `27e91d0582f40168aa218303c76e184bc78fa7a5d105aad0cfbad8458b441067`. Preserve the Apache-2.0 license and any notices in the assembled artifact. ## ONNX Runtime and conversion provenance The worker loads a CPU-only ONNX Runtime dynamic library from the extension package. Its upstream is Microsoft's [`microsoft/onnxruntime`](https://github.com/microsoft/onnxruntime) project, which publishes ONNX Runtime under the MIT license. The artifact lock pins official CPU-only ONNX Runtime 1.26.0 archives for each supported target. Assembly verifies the selected archive and library hashes and records the exact version, target, SHA-256, license, notices, and source location in the artifact manifest and SBOM. Read those records to learn which runtime a particular release shipped; the lock alone does not establish it. The reviewed OmniParser conversion used Python 3.12.13, PyTorch 2.8.0, torchvision 0.23.0, Ultralytics 8.3.199, ONNX 1.19.0, ONNX Runtime 1.22.1, opset 17, a fixed 1280 input, graph simplification, and no model-side NMS. These are conversion-provenance inputs; the Rust worker has no Python runtime dependency. The ONNX Runtime version used during conversion does not establish which runtime library was packaged for a release.