ffmpeg-skill Released: 42 Local Video Editing Tools for Coding Agents and MCP

Developer kajisho5 has released ffmpeg-skill, an open-source local video editing toolkit designed for coding agents and Model Context Protocol (MCP) clients. Operating entirely on local FFmpeg, ffprobe, and the Python standard library, the project requires no external cloud services, third-party accounts, or API keys.

Image source: https://github.com/kajisho5/ffmpeg-skill

When coding agents interact with video files, they frequently invent unsupported flags, trigger unneeded re-encodings, or report task completion without checking output files. ffmpeg-skill addresses these failure modes by enforcing a strict execution contract alongside deterministic local verification steps.

42 Local Editing Tools and Lossless-First Architecture

ffmpeg-skill bundles 42 granular CLI tools covering the complete post-production workflow. Capabilities range from basic cutting and joining to automated silence removal, aspect ratio and duration fitting, animated captioning, loudness normalization, and multicam synchronization.

• Structured Four-Stage Workflow: Every editing routine begins by measuring media streams with probe.py. Where feasible, the engine defaults to lossless stream copying (-c copy) before considering full re-encoding. Operations conclude with specification validation via check.py and visual contact sheet generation via verify.py.
• Typed CLI Arguments: The tools reject unstructured shell commands and raw filter graphs. All inputs are mapped through typed arguments derived from Python's standard argparse, eliminating command-line injection and syntax guessing.

Core 12 Tools by Default and Full 42-Tool MCP Server Integration

For agent integration, ffmpeg-skill supports Claude Code, Cursor, and Codex. Setup requires a single npx ffmpeg-skill command, with npx ffmpeg-skill doctor validating local installations of FFmpeg, ffprobe, and required media libraries.

• Context-Conscious MCP Surface: Under MCP server mode, tools/list exposes a curated core set of 12 tools by default to preserve the model's active context window. Operators can expose the complete 42-tool inventory by setting FFMPEG_SKILL_MCP_FULL=1. In addition, any tool remains callable by name via tools/call regardless of the listing mode.
• Self-Producing Execution Contract (SPEC): Tool schemas and CLI flags remain synchronized by deriving machine-readable JSON contracts (contract --json) directly from the underlying CLI parser at runtime.

Local Speech Engine Dependencies and Long-Render Timeouts

Deploying the toolset inside automated editing pipelines involves several operational constraints documented in the project contract.

• Local ASR Engine Requirements: Automated subtitle burning (caption.py --transcribe) and transcript-driven silence removal require an installed local speech model such as Whisper or Parakeet. When no local ASR engine is present, tools accept pre-existing transcript files instead of making remote calls.
• Rendering Timeout Ceilings: Operations operate under a default timeout ceiling of 1800 seconds (30 minutes). Long-form exports, such as full-hour high-resolution renders, require explicitly passing --timeout 0 or activating the --fast flag to prevent clean timeout termination.

Sources

• GitHub Repository: kajisho5/ffmpeg-skill