ffmpeg-skill MCP: a local video editor for your coding agent
ffmpeg-skill gives Claude Code, Cursor, Codex and other MCP clients 42 FFmpeg tools that cut, reframe, caption and check video on your own machine. The right MCP command and what I saw testing version 3.10.1.
ffmpeg-skill is an open source project from kajisho5 that turns FFmpeg into a set of tools a coding agent can call. You ask for something like a 60 second vertical cut with captions, and the agent probes the file, cuts it, reframes it, exports it for the platform and then checks the result, all on your machine. Nothing is uploaded and there is no API key.
It ships two ways: as an Agent Skill that drops scripts into your agent's skills folder, and as an MCP server for clients that cannot run shell commands. The repo is MIT licensed, had a little over 7,200 GitHub stars when I checked today, and showed up on both GitHub's daily trending page and appsgit's weekly trending MCP list, with about 1,240 stars gained between November 9 and November 16. I tested version 3.10.1, the latest release, on Linux.
What the tools do
There are 42 tools in total. By default the MCP server lists 12 core ones so your client does not load 42 schemas it rarely needs:
probe reads duration, codecs, frame rate, colour and audio streams.
cut keeps or removes time ranges, losslessly when keyframes allow and re-encoding when they do not.
fit changes duration or aspect ratio. export writes platform presets such as Reels.
silence finds and removes quiet parts. caption burns in or attaches subtitles, with speech to text from a local engine you install yourself.
loudness measures and normalizes audio. check verifies a file against a platform's limits. look writes a contact sheet so the agent can see the picture.
audio handles clean up, music with ducking and fades. graphics draws motion graphics templates. render builds a whole edit from one project file.
Set FFMPEG_SKILL_MCP_FULL to 1 to list all 42. Unlisted tools can still be called by name. probe, check and cropdetect are marked read only, and every other tool is marked destructive because it can overwrite a file if you allow it. The server also offers five prompts, reel, podcast, multicam, delivery_check and hdr, which return ready made step lists.
Install
You need FFmpeg 6.1 or later and Python 3.10 or later. The tools use only the Python standard library. Node 16 or later is needed only for the npx installer. On Debian or Ubuntu, FFmpeg is one apt install. On a Mac the README recommends the ffmpeg-full Homebrew formula, because the plain one lacks the subtitles, drawtext and zscale filters.
txsnippet
1sudo apt install ffmpeg2npx ffmpeg-skill doctor
Plain textUTF-8Ln 2, Col 24
doctor checks every FFmpeg encoder and filter the tools need and tells you, tool by tool, what will work on this machine.
Add the MCP server to Cursor
The mcp argument matters here. Running npx ffmpeg-skill on its own does not start a server. It installs the skill into your Claude Code skills folder and exits, which is what you want for the skill route, but an MCP client will just see a server that quits.
For Cursor and Codex skill folders the installer takes --cursor, --codex or --all.
What I saw when I tested it
I ran version 3.10.1 on Debian with the distribution's FFmpeg 7.1.5 and Python 3.13. I made a 12 second 720p test clip with a 440 Hz tone and three seconds of silence in the middle, then drove the server over stdio with npx -y ffmpeg-skill mcp.
doctor reported 86 capabilities available and nothing required missing. The only optional gaps were the speech engines for captions. It also warned that my FFmpeg 7.1.5 is older than the first upstream release with a fix for a recent decoder security bug, while saying a distribution package may have backported the fix.
The server started in under a second, listed 12 tools and 5 prompts, and 42 tools with FFMPEG_SKILL_MCP_FULL set.
probe returned the right facts: 12 seconds, 1280 by 720, 30 fps, H.264 with mono AAC.
silence with list found one silent stretch from 4.017 to 7.012 seconds, where I had put it, and proposed keeping the rest with a 0.2 second margin on each side.
cut with segments 0 to 3 and 8 to 12 wrote a 7.0 second file. It re-encoded because the second cut point was not on a keyframe, and it reported the reason and the nearest keyframes.
export with the reels preset made a 1080 by 1920 file, padding the 16 by 9 picture with bars by default. It warned that loudness was minus 21.8 LUFS against the preset's target of minus 14, and suggested the fix.
check on that file failed 1 of 13 Reels checks, loudness, and the call came back as an error with the full list. Exporting again with normalize gave minus 14.1 LUFS, and check then passed every row except a warning that there were no subtitles.
look wrote a 12 frame contact sheet with timestamps, so the agent can see what the picture looks like.
I did not test captions, since I had no speech engine installed, or GPU encoding.
Things to keep in mind
Releases come quickly, several in a single day this week. Pin a version in your MCP config, and rerun npx ffmpeg-skill to refresh skill copies, which do not update on their own.
Tools never change the input file, but they can replace an existing output when you pass overwrite. Point outputs at a scratch folder.
Export pads by default to reach a new aspect ratio. Ask for crop if you want the frame filled, and check the subject with look.
Your FFmpeg build decides what works. Run doctor first, especially on macOS.
Find it on PlugMCP
ffmpeg-skill is listed on PlugMCP with a link to the repo. If you want your own agent to search the directory for servers like this one, the PlugMCP directory MCP needs no key.