ffmpeg-skill
An Agent Skill and MCP server giving coding agents 42 local FFmpeg tools to cut, caption, reframe, normalize and export video, then verify the output.
At a Glance
ffmpeg-skill core under the MIT License, run locally with no cloud and no API keys.
Engagement
Available On
Listed Nov 2026
About ffmpeg-skill
ffmpeg-skill is an open-source Agent Skill from kajisho5 for Claude Code, Cursor, Codex and any agent that reads SKILL.md or speaks MCP. It teaches the agent a fixed workflow of probing a file, editing losslessly where possible, checking the result and verifying it, and ships 42 Python tools that drive local ffmpeg and ffprobe. Everything runs on the user's machine with no cloud and no API keys.
What It Is
The project is a local FFmpeg engine wrapped as structured, typed tools instead of shell strings. Each tool is a script with argparse flags covering cutting, joining, silence removal, fitting to duration or aspect ratio, captions and karaoke, overlays and motion graphics, HDR to SDR conversion, audio clean-up, sync, multicam, loudness, delivery checks, whole-edit project rendering and batch processing. Tools return structured JSON, and no tool accepts a caller-supplied filter graph or runs through a shell.
How the workflow works
Every job starts with probing the real file for duration, frame rate, resolution, colour and audio layout. The agent then edits, using stream copy where it can and re-encoding only when needed. Afterwards the output is probed again, checked against platform specs with check.py, and, when the picture changed, inspected as a contact sheet made by look.py. Tools support dry runs, saved plans, JSON output, timeouts and explicit overwrite consent, and never modify their inputs.
Contract and MCP surface
A machine-readable contract, available through contract --json, describes every tool's inputs, outputs, required FFmpeg capabilities and verification steps. The input schema is derived from each script's argparse parser. The MCP server builds its tool list from that contract, advertising a core set of 12 tools by default and all 42 when an environment variable is set. A doctor command reports which FFmpeg encoders, filters and fonts are present before a job fails.
Setup path
It installs with npx ffmpeg-skill into the Claude Code, Cursor or Codex skills folders, and is also offered as a plugin and extension for several agent clients. It requires FFmpeg 6.1 or later and Python 3.10 or later using only the standard library. Optional local speech transcription for captions and filler-word removal needs a separately installed engine such as faster-whisper.
Community Discussions
Be the first to start a conversation about ffmpeg-skill
Share your experience with ffmpeg-skill, ask questions, or help others learn from your insights.
Pricing
Open Source (MIT)
ffmpeg-skill core under the MIT License, run locally with no cloud and no API keys.
- MIT License
- 42 FFmpeg tools with machine-readable contract and MCP server
- No cloud · No API keys · Python standard library
Capabilities
Key Features
- 42 structured FFmpeg tools with typed arguments
- Lossless cutting with smart and frame-accurate modes
- Silence and filler-word removal
- Captions with karaoke animation and platform safe zones
- Reframe and fit to duration or aspect ratio
- HDR to SDR tone mapping and LUTs
- EBU R128 loudness normalization
- Platform export presets and delivery checks
- Declarative project rendering and batch processing
- Machine-readable contract via contract --json
- MCP server derived from the contract
- Dry-run plans and post-run verification
- Capability detection with doctor
- Multicam alignment and audio sync
