Reproduce the agent-authored edit demo

This demo turns a landscape clip using the dog demo's edit plan into an eight-second vertical cut. It is a deterministic replay of the JSON plan an LLM or MCP agent can author, so it does not need an API key, a model, or the [ai] extra.

Run it

Clone the repository, install FFmpeg, and provide a landscape video at least ten seconds long and at least 400×500 pixels. Run:

uv sync
uv run python examples/agent_edit_demo.py --source input.mp4 --output demo.mp4

The original dog source is a local test fixture and is not included in a clean checkout. --source replaces that default; --output selects the rendered file.

Give an agent the same brief

Turn the dog clip into a vertical social shot. Use the expressive middle section, increase the color slightly, add a restrained punch-in, and fade at both ends.

The replay uses this plan:

{
  "segments": [
    {
      "source": "src/tests/test_data/big_video.mp4",
      "start": 2.0,
      "end": 10.0,
      "operations": [
        {"op": "crop", "width": 400, "height": 500},
        {"op": "resize", "width": 512, "height": 640},
        {"op": "color_adjust", "contrast": 1.08, "saturation": 1.15},
        {"op": "punch_in", "zoom_factor": 1.04, "attack_frames": 12, "release_frames": 12},
        {"op": "fade", "mode": "in_out", "duration": 0.35}
      ]
    }
  ]
}

VideoEdit.from_dict() parses the same JSON shape exposed to an LLM. The script calls validate() before rendering, then streams the edit to the output file. To let an agent select scenes by catalog id instead of receiving exact time bounds, use the MCP workflow.