Skip to content

← The Lab

ffmpeg-skill hits GitHub Trending again: what changed in the 3.x line of the local video editor for coding agents

The free, MIT-licensed skill gives Claude Code, Codex and Cursor 42 FFmpeg tools with typed arguments plus an MCP server, all running on your own machine. It's back on GitHub Trending at about 475 stars a day and passed 6,600 stars. The 3.x line it shipped since our October write-up is worth a look if you tried it before.

What it is. Agents 'know' FFmpeg, but they still guess frame rates, re-encode a whole file for a quick trim, and call a job done without checking. ffmpeg-skill hands the agent 42 tools instead: cut, join, captions, reframe, loudness, stabilize, platform exports and more. Each tool has a machine-readable contract (npx ffmpeg-skill contract --json), so the agent reads the exact flags rather than inventing them. The loop is probe, edit, verify: measure the file first, make the edit, then check the output against what was asked. No cloud, no account, no API key. You only need FFmpeg 6.1+ and Python 3.10+.

Install. Run npx ffmpeg-skill for Claude Code, add --cursor or --codex for those apps (or --all), and npx ffmpeg-skill mcp to run it as an MCP server. npx ffmpeg-skill doctor tells you what's missing. There's also a Claude Code plugin through the repo's marketplace. Re-run the npx command to update, because copies don't refresh themselves.

What 3.x changed. v3.0.0 (Oct 31) dropped FFmpeg 5.x and Python 3.9, and every tool now refuses an older build up front, with install steps, instead of failing halfway. It also changed some cut.py defaults. Stream-copy cuts now keep the MP4 edit list, so picture and sound start exactly at --start. SDR HEVC sources re-encode as HEVC rather than x264. The variable-frame-rate guard now judges by measured packet timing. Each change has a flag to restore the 2.x behavior, and the contract lists what was removed. probe now reports source timecode as well. Later releases added DNxHR presets, alpha masters and key/fill exports, loudness targets (streaming, EBU R128, Japanese broadcast) and title/action-safe guides (3.2). They added a source_carry field that says what an output lost from its source, plus Codex and Gemini CLI plugin installs (3.3), sentence-aware silence margins for talking-head cuts (3.5–3.6), and tempo and sample-rate controls in audio.py (3.7). v3.10 (Nov 15) makes a multi-clip render and its exported FCPXML, EDL or OTIO timeline agree on frame size and fps, so a timeline you hand to Premiere, Resolve or Final Cut matches the MP4.

The numbers are the maintainer's. The README reports 92/92 verification steps passed on a 10-file real-device corpus (GoPro, DJI, iPhone including Dolby Vision, HDR10, Android screen recordings). Audio sync detection landed within 10 ms in 40 of 40 cases, silence detection missed no gaps across 20 known cases, and scene detection scored F1 0.97 on 53 hard cuts. The README is upfront that the full corpus hasn't been run on Windows yet. Treat these as the project's own Linux and macOS measurements.

Where it fits. It edits footage you already have. It's not a renderer like Remotion or HyperFrames, but a file either of them writes is a normal input. The same author's remotion-ffmpeg-skill joins the two: it measures footage with ffmpeg-skill, draws Remotion graphics from those measurements, then checks and delivers the result. It also keeps showing up as the cutting layer in community stacks. Today @plusxbt put it at #2 in a 14-repo 'video editing room for Claude Code', between video-use and Claude Eyes.

Signals: on GitHub Trending (Python #14 on Nov 13, then all-languages #14 and Python #5 on Nov 16 at about 475 stars a day), and the tracked curator @RoundtableSpace posted it on Nov 16 at 3:45 PM MST. Trending plus a curator pick within 72 hours makes it HOT by our rule.

Skills & repos in our directory

More from the Lab

Topics: Remotion · Video MCP servers