Your Coding Agent Can Edit Video Now—And Check Its Own Work
You've probably handed a coding agent a tedious refactor or a test suite to fix. But what happens when the task isn't code? Say you want to cut the dead air out of a recording, reframe a 16:9 clip to vertical, or normalize the loudness of an audio track. An agent that can write code is only half the story—it still needs a way to actually touch the footage and confirm the result. ffmpeg-skill gives it exactly that: a local video editor built on FFmpeg, plus a workflow for checking its own output.
What It Does
ffmpeg-skill is an Agent Skill for Claude Code, Cursor, Codex, and any agent that reads a SKILL.md file. It teaches the agent a fixed workflow—probe, edit losslessly where possible, check, verify—and ships 42 tools that do the real work through ffmpeg and ffprobe. Those tools cover cutting, joining, silence removal, fitting to a target duration or aspect ratio, captions and karaoke, overlays and motion graphics, HDR-to-SDR conversion and LUTs, audio cleanup and dynamics, sync with drift correction, multicam, loudness, delivery checks, whole-edit project rendering, and batch folder processing.
Everything runs locally. No cloud, no API keys, and the tooling sticks to the Python standard library. The same 42 tools are also callable as MCP tools, so any MCP client can use them. By default, tools/list advertises a core set of 12 to keep the client's context small. Installation is a single command: npx ffmpeg-skill.
Why It's Cool
The agent verifies its own work. This is the part that stands out. The workflow isn't just "run an FFmpeg command and hope." It's probe, edit, check, verify—so the agent measures the output file and confirms it matches what was asked for. The README's own example shows a finished file measured at 59.98 seconds, 1080x1920, 30 fps. That feedback loop is what turns a video editor into something an agent can actually use reliably.
Lossless first. The workflow explicitly prefers lossless edits where possible. That's a sensible default for agents, which are prone to re-encoding things they didn't need to touch and quietly degrading quality along the way.
Real before/after demos you can rebuild. The project documents its features with side-by-side demos—input on the left, output on the right—each with the exact command underneath. The animated captions with karaoke highlighting, the 16:9-to-9:16 crop reframe, silence removal, and loudness normalization are all shown this way. Better still, every demo is built by python3 demos/build.py, so you can regenerate each frame yourself. The footage is synthetic (FFmpeg's own test sources) except for three demos that use AI-generated clips of people who don't exist.
Broad compatibility, honestly tested. The CI tests against FFmpeg 6.1, 7.1, 8, and 9, which tells you the project takes version drift seriously rather than assuming one install.
Context-aware MCP defaults. Advertising only 12 core tools by default is a small but thoughtful decision. Agent context is precious, and flooding a client with 42 tool definitions up front isn't helpful. The full set is still there when you need it.
How to Try It
- Make sure you have FFmpeg installed locally (the tools shell out to
ffmpegandffprobe). - Run the installer:
npx ffmpeg-skill
- Point your agent at it—Claude Code, Cursor, Codex, or any client that reads
SKILL.md. MCP clients can connect to the tool server directly. - Start with something small. Ask the agent to probe a file, then try a simple cut or a silence removal, and watch it verify the result.
The demos page is the best place to see what each tool does before you commit to a workflow, and the repository has the full tool list and MCP details:
https://github.com/kajisho5/ffmpeg-skill
Final Thoughts
ffmpeg-skill is aimed at people who already use coding agents and want them to handle video chores without leaving the terminal or uploading footage anywhere. If you're comfortable with FFmpeg, you'll recognize the underlying power; if you're not, the skill layer is what makes it approachable. The self-verification loop is the real value here—it's the difference between an agent that tries something and an agent that knows it worked. Worth a look if your workflow keeps brushing up against video files.