AI Agents Can Now Edit Video, and the First Thing the Comments Ask Is Whether They'll Cut It Wrong

npx ffmpeg-skill. This command adds 42 video editing tools to Claude Code, Cursor, Codex, or any MCP client. Editing runs through local FFmpeg, and when it finishes it can check its own output. Project: https://github.com/kajisho5/ffmpeg-skill . Original tweet: https://x.com/tom_doerr/status/2115405924646985924 .

FFmpeg is an old standby in video processing; transcoding, trimming, merging, and extracting audio all rely on it. The problem is that the command-line parameters are numerous and finicky: if a filter graph has one ordering mistake, the output can quietly break. For many people, their memory of it is that they search for flags every time they use it. One commenter said ffmpeg flags are something he has had to Google every time for years. Wrapping common workflows into named tools so the agent first chooses a tool and then fills in parameters is not a strange direction.

The project description lists 42 tools, covering cutting, merging, noise reduction, fitting duration and aspect ratio, adding titles and karaoke lyrics, overlays and motion graphics, HDR to SDR, audio cleanup, loudness, delivery checks, full project rendering, and batch processing folders. All tools can also be called as MCP tools; by default it advertises only 12, with the goal of not blowing up the client context.

The comments are more worth reading than the README. Someone asked whether the verify step really rejects bad exports. Someone else complained that 42 tools are too many, saying 5 that always work would be enough. PuneetSingh warned to point the agent at a copy of the footage folder, not the original footage, because when the agent runs ffmpeg it may overwrite files with the same name in one step. He also asked whether the 42 tools are loaded on every request, which would eat into the context the model has left for actual editing. scanner's judgment was more specific: the model knows the CLI, but filter graphs and flag order are where it tends to make quiet mistakes. He cared about whether the agent can get progress and exit codes during long renders, or only gets the final file. Ricardo Dias focused on trim's defaults: with -c copy, FFmpeg can only cut on keyframes, so the start of the clip may be off by a second or two; only re-encoding is frame-accurate, but slow. He wanted to know which option the tool chooses when the agent just says trim. Guillermo said FFmpeg has long since stopped being the hardest part; choosing which take to use is the editing. These responses pull the question back from tool count to defaults, verification, and safety boundaries.

42 tools are capability, but also a burden of choice. For an agent, the longer the menu, the larger the surface area for choosing wrong. What is really worth examining is whether it writes FFmpeg's pitfalls into the tool contracts, whether the default parameters are safe, whether verification can reject bad output, whether the rendering process is observable, and whether the context budget is controlled. Do these well, and 42 is not too many; do them poorly, and even 5 can make an agent cut the wrong clip. For ordinary users, remember one thing first: have the agent work on a copy of the footage. For developers, read SKILL.md and the tool definitions first, then see how it verifies exports. The project is already open source under the MIT license; the rest is left to actually running it once.