Clip videos locally, with offline processing.

ClipMill processes the default video workflow locally: source inspection, speech transcription, clip suggestions, editing, visual checks, and export. A local recording does not have to be uploaded for that workflow.

Local processing is useful when a recording is large, when an upload would be slow, or when you want the source footage to stay on your disk. It also means your available memory and processing hardware affect how quickly a job finishes.

The download page has builds for Apple silicon Macs, Windows x64 in beta, and Linux x86_64. Check the platform requirements before installing.

First launch opens Set up ClipMill, which lists the processing components and recommended models to download. Components include the programs that run the models; model weights are separate files. Setup verifies downloaded files against their expected hashes.

The local editorial model is Qwen3.5. Apple silicon uses MLX. Windows and Linux use a bundled llama.cpp server, with a suitable Vulkan graphics card when available or the CPU otherwise. Models offers a smaller 4B option as well as 9B on Windows and Linux. CPU-only analysis can take many minutes.

Use the model choices and memory information in the app instead of assuming every computer can run every model comfortably. Without an editorial model, the built-in picker can still propose moments.

After components and models are installed, import a video already on your computer and run local analysis. Review the clips, adjust captions and framing, and export the finished files without uploading the source.

If you want the app to avoid its scheduled release look-up, turn off Say when a new version is out in Settings → About. The current app can check GitHub once a day for update notices. It does not automatically install a new app version.

Starting another model download, looking up a model repository, or importing a YouTube video needs a connection. Those actions are separate from local analysis. See installation and updates for the setup details.

YouTube import fetches the video you select before the local workflow begins. Downloading a model or component contacts the service hosting those files. The app's update notice contacts GitHub when enabled.

Optional cloud editorial processing is separately enabled. An analysis must opt in, select a supported provider and model, and supply a spending limit. Cloud proposals and reviews send transcript context; source frames and visual checks remain local. There is no automatic switch from local analysis to a cloud provider after a failure.

These differences matter when choosing a workflow: local analysis uses your hardware, while optional provider processing can have separate usage charges.

Local Lock reports whether a network-enabled operation has started in the current background-service session. Its counter records operation starts, including work fulfilled from a cache. It does not count bytes or prove that a remote request completed.

The app checks task permissions, but the indicator is not an operating-system firewall or a sandbox for every desktop process. For the exact scope, read privacy and local processing.

Projects, components, models, and logs stay in the local application data folder. Settings and storage explains where to find them and how to manage disk use.