Nock CLI

Open models that live on your machine. Ask questions about your videos, offline.

nock video summarize my_video.mp4
nock video ask my_video.mp4 "What error message shows on screen?"
nock video find my_video.mp4 "when the error message shows up"
nock video summarize "https://www.youtube.com/watch?v=..."

Everything runs on your computer. Frames are read into a temporary folder, sent to a model running on 127.0.0.1, and deleted when the answer is ready. Nothing is uploaded.

Install

You need Python 3.9 or newer and ffmpeg.

pipx install nock-cli
nock setup
nock doctor

nock setup downloads, once:

• the engine: an official llama.cpp build (about 35 MB; Vulkan on Windows and Linux, Metal on Apple Silicon). Already have llama-server? Nock uses it.
• the model: Nock Video 4B (about 3.4 GB).

Files go to ~/.nock (set NOCK_HOME to use another folder).

Links

Links are downloaded to your own machine with yt-dlp, which you install yourself:

pipx inject nock-cli yt-dlp

Only use videos you have the right to use.

What to expect

Nock Video 4B is a preview. On our test of 20 videos, through llama.cpp at 64 frames, it found the moment in 42 of 56 questions, answered 56 of 60 questions about what is on screen, and gave 6 of 7 good summaries. Full results, including the misses, are on the model card.

• It reads frames only, not sound.
• A long video is read as at most 64 frames, so short events in long videos can be missed.
• 64 frames used about 5.5 GB of GPU memory in our test. On GPUs under 6 GB, Nock uses 32 frames (finding the moment drops: 42 to 35 of 56 on our test). --frames overrides it.
• Without a GPU it works, but slowly: minutes per question.

Options

Server logs are in ~/.nock/logs/ if something goes wrong.

License

Apache-2.0. The model has its own license on its model card (Apache-2.0, from Qwen3.5 4B).