clipwright-stabilize

MCP tool for video shake detection and OTIO timeline stabilize annotation generation.

Overview

Runs ffmpeg vidstabdetect to generate a .trf transform file, estimates shake severity from the binary TRF1 data (best-effort heuristic), and writes a stabilize directive to timeline-level metadata["clipwright"]["stabilize"].

Performs detection only (OTIO annotation); realization (vidstabtransform application) is done once by clipwright-render (design M3: separation of detection and application).

Severity estimation:

• Reads the binary TRF1 file produced by vidstabdetect.
• Scans all IEEE-754 little-endian doubles, computes mean absolute value.
• Normalises by a pinned heuristic constant (_NORM_PX = 30.0 px) to derive a severity in [0.0, 1.0].
• Returns severity=null when the file cannot be parsed (non-fatal; render does not use severity).

Prerequisites

• Python 3.11 or later
• ffmpeg compiled with --enable-libvidstab must exist on PATH or full path set in environment variable CLIPWRIGHT_FFMPEG. Standard distribution builds (apt, brew, choco) may NOT include libvidstab. Use a build that explicitly enables the vidstab filter.

export CLIPWRIGHT_FFMPEG=/path/to/ffmpeg-with-libvidstab
export CLIPWRIGHT_FFPROBE=/path/to/ffprobe

MCP Tool

clipwright_detect_shake

Parameters

Return value

The tool returns a ToolResult envelope:

{
  "ok": true,
  "summary": "Shake analysis of video.mp4 complete. severity=0.312, shakiness=5, smoothing=30. Stabilize directive and video.stabilize.trf written; apply with clipwright-render.",
  "data": {
    "severity": 0.312,
    "shakiness": 5,
    "accuracy": 15,
    "smoothing": 30,
    "trf_basename": "video.stabilize.trf"
  },
  "artifacts": [
    {"role": "timeline", "path": "out.otio", "format": "otio"},
    {"role": "analysis", "path": "video.stabilize.trf", "format": "trf"}
  ],
  "warnings": []
}

When libvidstab is not compiled into the ffmpeg build, the tool returns UNSUPPORTED_OPERATION with installation guidance (no path or raw stderr exposed).

Dependencies

ffmpeg is invoked as a separate process (via PATH or environment variable) for license independence.

Detection and Render Two-Phase Flow

1. detect (this tool): ffmpeg -i <media> -vf "vidstabdetect=result=<stem>.stabilize.trf:shakiness=<n>:accuracy=<n>" -f null - generates .trf and saves the stabilize directive to OTIO annotation.
2. render (clipwright-render): reads metadata["clipwright"]["stabilize"] and applies vidstabtransform=input=<basename>:smoothing=<n> in the ffmpeg filter graph using cwd=<trf parent directory> for Windows-safe relative path resolution.

Installation and Startup

Within a uv workspace:

uv run --package clipwright-stabilize clipwright-stabilize

Or install directly:

uv add clipwright-stabilize
clipwright-stabilize