ffmpeg-skill gives Claude Code and Cursor 42 local video-editing tools
Ask an AI agent to keep 0:45–1:20 of a recording and turn it into a vertical Reel, and it will often write an FFmpeg command by guesswork, then report "done" without ever opening the file. ffmpeg-skill (331 GitHub stars in a day) replaces that guessing with 42 ready-made local tools for Claude Code, Cursor and Codex: the agent measures the clip, edits it and checks the result. There is no cloud and no API keys, and the video stays on your machine.
AI-processed from GitHub Trending AI; edited by Hamidun News
Imagine this task: you have an hour-long recording of a talk, and you want to keep the segments 0:45–1:20 and 5:00–5:25 and turn them into a one-minute vertical clip for Reels. Hand that to an AI agent and it usually writes an FFmpeg command (FFmpeg is a popular command-line video processing tool) by guesswork. It assumes a frame rate, picks a codec the container doesn't support, re-encodes what it could simply have copied, and reports "done" without opening the result.
The ffmpeg-skill repository, which gained 331 stars in a day on GitHub trending, removes that guessing. It gives Claude Code, Cursor, Codex and any MCP client (MCP is the protocol agents use to call external tools) 42 ready-made local editing tools. The agent doesn't write commands itself.
It calls a tool with clear parameters and checks what came out. Everything runs on your computer, with no cloud and no API keys.
What the 42 tools can do
The tools run on top of a local FFmpeg and are written in the Python standard library, with no third-party dependencies. The set includes cutting and joining, silence removal, fitting to a target duration and aspect ratio, subtitles (including word-by-word, karaoke-style ones), titles and graphics overlaid on video, HDR-to-SDR conversion, audio cleanup, loudness normalization, sync with drift correction, multicam and batch processing of folders.
There are templates for popular platforms: they build a clip for seven video platforms. For TikTok, for example, the template crops to 9:16, animates the subtitle words, avoids the app's interface zones and sets loudness to −14 LUFS (a standard loudness unit for streaming). An example command: render.py talk.mp4 --template tiktok --cues cues.txt. The subtitle text is taken as is; the tool does not rewrite it.
How the agent is kept from guessing
The agent follows a fixed order: first measure the file, then edit, losslessly where possible, then verify and confirm the result. Decisions rest on the measured duration, fps, resolution, color and audio tracks, not on the file name. Re-encoding kicks in only when there is no way around it: for frame-accurate cuts, filters or a format change. In all other cases the streams are simply copied.
The agent passes named parameters, such as --aspect 9:16 --fit crop or --threshold -35, while the tool itself builds the command and the filter graph. It does not accept an arbitrary graph from the caller, and nothing is run through a shell.
After the job, the tool checks the file it wrote: the actual duration, frame size and audio. In the repository's example, the agent reports: "final.mp4: 60.00 s, 1080×1920, 30 fps, AAC stereo" and attaches a contact sheet of frames that it has inspected. The "looking" here is done by the multimodal agent itself: the repository has no face recognition or "interesting moment" detection. What to cut and what counts as a good result is up to the agent or the human.
Source protection and the contract
A destination file that matches the input is rejected, and replacing an existing file requires explicit consent. Dry-run mode prints the commands and writes nothing. A plan can be saved and run later, and it refuses to run if the input file has changed. A hung ffmpeg process is forcibly terminated, and the half-written file is deleted.
The most interesting part is the machine-readable contract. For each tool, it describes the inputs, outputs, required FFmpeg components and the way the result is verified. The MCP schema is built from the same argument parsers as the command line, and CI fails if the two diverge. As a result, the description can't go stale. To avoid bloating the client's context, MCP shows 12 core tools by default; the full list of 42 is available separately, and any tool can be called by name.
Limitations
You need FFmpeg installed with the required components. The doctor command shows what is missing: without libvidstab, for example, some tools will be unavailable. Speech recognition is optional: it works if Parakeet or Whisper is installed, and without them the tools accept a ready-made transcript. Model weights are downloaded on first run, but no audio leaves the machine. Copies of the skill are not updated automatically. In his own position article on verifying results, the author rated ffmpeg-skill as only partially meeting the discipline he describes.
What to do with it
Agent developers should borrow the approach itself: narrow, typed tools, a contract from which MCP is derived, and mandatory verification of the result after every operation. The same technique works for documents, images and data. For people who simply need to edit videos, the project offers a ready foundation for local automation, with no subscriptions and no uploading of files to the cloud.
- Nov 18, 2026The Best Coding Agent Controls a Robot in 77% of Tasks, but With Sensors Alone Only 30%
- Nov 18, 2026Moonshot AI Shifts Kimi From Code to Science: Gene Analysis in 90 Minutes Instead of 4 Hours
- Nov 18, 2026Duco cuts agent context by 90% and onboards bank reconciliations 10x faster
- Nov 18, 2026Epic to remove Blueprints from Unreal Engine 6: AI works with code, not diagrams
Key Questions Answered
What the 42 tools can do?
The tools run on top of a local FFmpeg and are written in the Python standard library, with no third-party dependencies. The set includes cutting and joining, silence removal, fitting to a target duration and aspect ratio, subtitles (including word-by-word, karaoke-style ones), titles and graphics overlaid on video, HDR-to-SDR conversion, audio cleanup, loudness normalization, sync with drift correction, multicam and batch processing of folders.
How the agent is kept from guessing?
The agent follows a fixed order: first measure the file, then edit, losslessly where possible, then verify and confirm the result. Decisions rest on the measured duration, fps, resolution, color and audio tracks, not on the file name. Re-encoding kicks in only when there is no way around it: for frame-accurate cuts, filters or a format change. In all other cases the streams are simply copied.
What are the key details on Source protection and the contract?
A destination file that matches the input is rejected, and replacing an existing file requires explicit consent. Dry-run mode prints the commands and writes nothing. A plan can be saved and run later, and it refuses to run if the input file has changed. A hung ffmpeg process is forcibly terminated, and the half-written file is deleted. The most interesting part is the machine-readable contract. For each tool, it describes the inputs, outputs, required FFmpeg components and the way the result is verified.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.