kinescript

Checked without watching.

An agent cannot watch a video. Each command returns something it can read: one image of many frames, a JSON list of problems, the box of every layer at a given time.

kine look card-flip.json --boxes

Frames spread over the scene in one image, each with its time and frame number. With --boxes every layer is outlined and the top-level ones are named.

A contact sheet of eight frames from the card flip demo. Each frame is labelled with its time and frame number, and each layer is outlined in green with its id.
The actual output for the card flip demo, eight frames. Open the image for full size.

kine look explainer.json --trails

Each moving layer's path drawn over the frames, a dot for every frame, so their spacing shows where it eases, overshoots or stops. --onion 3 draws the three frames before each one over it, and kine probe --curve title.x gives a key's value and speed at every frame as JSON.

A contact sheet of eight frames from the explainer demo, with the path of each moving layer drawn over every frame as a line of coloured dots, one a frame.
The actual output for the explainer demo, eight frames. Open the image for full size.

kine check problems.json

Every problem at once, errors first, each with its layer, time range and a fix. Contrast is measured against the pixels actually behind the text. When the scene has sound it mixes and measures it too, and reports a voice cut off or masked by the music, an effect that lands late, and a mix off its loudness target. The command exits with status 1 while there are errors.

problems.jsonfour problems
{  "canvas": {"w": 1920, "h": 1080, "fps": 30, "bg": "#f2f0e8"},  "layers": [    {"id": "panel", "type": "rect", "x": "50%", "y": "62%", "w": 1300, "h": 260, "radius": 20, "fill": "#ff633f"},    {"id": "title", "type": "text", "text": "Quarterly results", "x": "50%", "y": "6%", "size": 120,     "fill": "#0a0c10", "dur": 4},    {"id": "label", "type": "text", "text": "Orders up 18 percent on last year", "x": "50%", "y": "62%",     "size": 60, "fill": "#ff9a80", "dur": 1.2},    {"id": "note", "type": "text", "text": "Source: annual report", "x": "50%", "y": "88%", "fontSize": 40,     "fill": "#4a4f5c", "dur": 4}  ]}
output
{
  "ok": false,
  "errors": 1,
  "warnings": 2,
  "issues": [
    {
      "level": "error",
      "code": "text-offscreen",
      "layer": "title",
      "from": 0.0,
      "to": 4.0,
      "message": "text box [469.7, -6.0, 1450.3, 135.6] leaves the 1920x1080 canvas",
      "fix": "move it inward, lower size, or set maxWidth"
    },
    {
      "level": "warning",
      "code": "read-time",
      "layer": "label",
      "from": 0.0,
      "to": 1.2,
      "message": "7 words on screen for 1.20s; viewers need ~2.5s",
      "fix": "extend dur or cut words"
    },
    {
      "level": "warning",
      "code": "contrast",
      "layer": "label",
      "from": 0.72,
      "to": 0.72,
      "message": "contrast against what is behind it is 1.4:1 (want 4.5:1); it overlaps 'panel'",
      "fix": "move it clear of 'panel', or add a box or a shadow"
    },
    {
      "level": "info",
      "code": "renamed-key",
      "layer": "note",
      "message": "'fontSize' read as 'size'",
      "fix": "write 'size'"
    }
  ],
  "duration": 4.0
}

kine probe card-flip.json --at 1.5

Each layer's start and end, and at a given time, where every visible layer is on screen in pixels. Trimmed here to four layers: the title, the back of the first card, which has turned past halfway, and the text on the second card.

{
  "at": 1.5,
  "visible": {
    "title": {
      "rect": [467, 158, 1453, 252],
      "alpha": 1.0,
      "type": "text",
      "scale": 1.0,
      "text": "Change a line. Flip a card."
    },
    "card-1-back": {
      "rect": [235, 464, 657, 762],
      "alpha": 1.0,
      "type": "group",
      "scale": 0.9
    },
    "card-1-new": {
      "rect": [292, 570, 598, 651],
      "alpha": 1.0,
      "type": "text",
      "scale": 0.9,
      "text": "+ \"dur\": 4.5"
    },
    "card-2-old": {
      "rect": [774, 540, 1134, 614],
      "alpha": 1.0,
      "type": "text",
      "scale": 1.1,
      "text": "- \"text\": \"Q3\""
    }
  }
}
While it is written
kine watch scene.json makes the contact sheet again each time the file is saved, so an agent or a person keeps one image open beside the scene.
More overlays
--onion 3 draws the three frames before each one over it, fainter the older they are. --wire draws every model's triangles; --depth shows each 3D stage's depth image. --at 0.5 2 end picks the times instead of even spacing.
One key over time
kine probe scene.json --curve title.x prints a key's value at every frame and how fast it changes, where it holds still and its fastest frame.
Brand rules
In a branded scene kine check also reports text and chart colours below the contrast they need, a logo drawn too small, and a scene overriding a part the brand locks. --brand-strict adds colours and fonts the brand does not have, and text outside its safe area.

MCP

kine mcp serves the same tools to MCP clients over stdio. Tools take a scene as a file path or as the JSON object itself, and the look and still tools return the image inline.

MCP client config
{
  "mcpServers": {
    "kinescript": {"command": "kine", "args": ["mcp"]}
  }
}
  • kine_docsthe scene reference
  • kine_checkproblems as JSON
  • kine_looka contact sheet, as an image
  • kine_stillone frame, as an image
  • kine_probetimeline and layer boxes
  • kine_rendera draft, a range or the final file
  • kine_voicespeaks a scene's text and times its words
  • kine_feedbackreads and answers the notes left in the editor
  • kine_brandthe brand library, everything the brand wizard does

kine docs prints the scene reference, about 750 lines: rules, a complete scene, every layer type with its defaults, and recipes. It is written to fit in a model's context. kine docs 3d, avatars, brands and explainers print the topics kept out of it.