ffmpeg-skill Gives AI Coding Agents a Verifiable Video Workflow With 42 Tools

A new open-source project, ffmpeg-skill, packages 42 Python scripts around FFmpeg, the command-line media engine. It exposes those tools to Claude Code, Cursor, Codex, and Model Context Protocol clients, allowing an agent to inspect media, edit it, verify the output, and return structured JSON. The repository crossed 3,600 GitHub stars within weeks and appeared on GitHub Trending.

Coding agents often generate plausible FFmpeg commands while making operational mistakes: assuming the wrong frame rate, selecting a codec the container cannot hold, re-encoding streams unnecessarily, or declaring success without checking the resulting file. ffmpeg-skill addresses those failures with a prescribed sequence that probes the source, preserves streams where possible, checks the output, and creates a contact sheet for visual inspection.

All processing stays on the local machine. The tools require ffmpeg and python3 on the system path, while installation through the npm package requires npx. The Python scripts use the standard library and require no cloud service or API key, which supports offline work and footage that cannot be uploaded.

Probe, edit, inspect

Each job follows a workflow designed to give the agent reliable information before and after an edit:

1. Probe the source. probe.py reads duration, frame rate, resolution, codecs, audio streams, and HDR metadata.
2. Plan the operation. The agent selects tools and can inspect the generated FFmpeg commands before changing files.
3. Edit conservatively. The scripts use stream copying when the requested cut points and output format allow it.
4. Check the file. check.py re-probes the output and evaluates it against the selected delivery target.
5. Inspect the picture. look.py generates a contact sheet that a multimodal agent can review for framing, captions, and visible errors.

Installation takes one command