FFmpeg MCP Server

ffmpeg-mcp-server

English | 한국어

A Model Context Protocol (MCP) server that gives AI tools access to FFmpeg and FFprobe to process video and audio locally.

Supports cross-platform GPU hardware encoding (Apple Silicon, NVIDIA NVENC, Intel QuickSync, AMD AMF), live progress updates via MCP progress tokens, instant cancellation via AbortSignal, and inlines image previews directly for multimodal vision models.

---

Why This Exists

When AI agents process media, they often hallucinate invalid flags, syntax errors, or incompatible pixel formats.

This server provides 19 typed tools using Zod schemas. The AI passes parameters (such as { inputPath: "video.mov", preset: "web_mp4" }), and the server runs the valid FFmpeg command locally with proper error handling, hardware acceleration, and progress notifications.

---

Key Features

• Universal Hardware Acceleration: Automatically detects and leverages hardware encoders across macOS (VideoToolbox), Linux/Windows GPUs (NVIDIA NVENC, Intel QuickSync, AMD AMF, and VA-API) for 5x–20x faster exports with low CPU usage.
• Real-time Progress Reporting: Emits live percentage updates via standard MCP progress tokens (notifications/progress), enabling progress bars in modern AI clients.
• Instant Cancellation: Listens to client AbortSignal events to immediately terminate child FFmpeg processes when a generation is stopped.
• Multimodal Vision Previews: Inlines base64 image data for thumbnails, audio waveforms, and multi-frame contact sheets so vision models (Claude 3.7, GPT-4o, Gemini) can "see" media content.
• Social Formats & Automation: Reformats landscape video to 9:16 vertical video with blurred backgrounds, burns customizable hardcoded subtitles, and stacks top/bottom split-screen reels for TikTok and Instagram Reels.
• Audio Engineering: Normalizes loudness to broadcast/streaming standards (EBU R128 / -16 LUFS), cleans vocal audio with 80Hz rumble removal, and trims dead air while preserving lip-sync.
• Scene Intelligence & Multiplexing: Detects visual scene changes and cut points for smart editing workflows, and combines or replaces video/audio tracks with zero-loss stream copy.

---

Included Tools (19 Total)

---

MCP Resources & Prompts

Resources

• `media://capabilities`: Inspects local FFmpeg installation, active version, detected hardware accelerators (VideoToolbox, NVENC, QSV, AMF, VAAPI), and supported video encoders.
• `media://presets`: Standard reference table of social aspect ratios, platform upload file size limits, and broadcast audio loudness targets.

Prompts

• `social_clip_generator`: Interactive workflow that turns raw landscape video into vertical shorts with blurred background padding, trimmed silence, and hardcoded captions.
• `whisper_audio_prep`: Standard pipeline for extracting 16kHz mono WAV audio and normalizing loudness for speech-to-text models.

---

Requirements

• Node.js >= 18.0.0
• FFmpeg & FFprobe installed locally:
• macOS: brew install ffmpeg
• Linux (Ubuntu/Debian): sudo apt update && sudo apt install ffmpeg libass-dev
• Windows: winget install Gyan.FFmpeg or choco install ffmpeg

---

Installation

Quick Start (Claude Desktop, Cursor, Antigravity)

Run directly with npx (no manual cloning or building required):

{

"mcpServers": { "ffmpeg": { "command": "npx", "args": ["-y", "@ash_j1997/ffmpeg-mcp-server"] } } }

Configuration File Locations:

• Claude Desktop (macOS): ~/Library/Application Support/Claude/claude_desktop_config.json
• Claude Desktop (Windows): %APPDATA%\Claude\claude_desktop_config.json
• Antigravity IDE: ~/.gemini/config/mcp_config.json
• Cursor / Windsurf: Add as a stdio server with command npx and args ["-y", "@ash_j1997/ffmpeg-mcp-server"]

---

Local Development & Testing

1. Install dependencies:

   npm install

1. Run build and automated test suite:

   npm test

---

Acknowledgements & Community

Special thanks to @rajarshidattapy for suggesting the stream multiplexing (mux_media) and scene transition detection (detect_scenes) capabilities to expand the server beyond format conversion.

---

License

MIT © Ashutos1997