transcribe

transcribe

A Claude Code plugin that transcribes audio and video files into organized markdown documents with speaker diarization.

Features

• Transcribes any audio format (m4a, mp3, wav, ogg, flac, etc.)
• Transcribes any video format (mp4, mkv, mov, avi, webm, etc.) — extracts audio automatically
• Speaker diarization (identifies who said what)
• Three output formats: Summary + Transcript, Summary only, or Transcript only
• Guided first-time setup — walks you through everything below
• Supports local or remote (SSH) server deployment

Prerequisites

On your local machine (where Claude Code runs)

Dependency	Purpose	Install
Claude Code	Plugin host	claude.com/claude-code
ffmpeg / ffprobe	Video-to-audio extraction, duration probing	sudo apt install ffmpeg (Debian/Ubuntu) or brew install ffmpeg (macOS)
Python 3.8+	Only needed if running the MCP server locally	sudo apt install python3 python3-venv
SSH client	Only needed if the MCP server is on a remote host	Pre-installed on most systems

On the MCP server host (local or remote)

Dependency	Purpose	Install
Python 3.8+	Runs the transcription server	sudo apt install python3 python3-venv python3-pip
ffmpeg	Converts audio to MP3 before uploading to AssemblyAI	sudo apt install ffmpeg
AssemblyAI API key	Speech-to-text engine	Free account at assemblyai.com

Installation

# Add the marketplace (one-time)
claude plugin marketplace add https://github.com/3D-Stories/transcribe.git

# Install the plugin
claude plugin install transcribe@transcribe

To update later:

claude plugin update transcribe@transcribe

To reinstall from scratch:

claude plugin remove transcribe@transcribe
claude plugin install transcribe@transcribe

Usage

/transcribe path/to/recording.mp4
/transcribe meeting-notes.m4a

On first run, the skill walks you through setup (see Architecture for what gets installed where).

Output

Transcripts are written to a transcripts/ directory in your project:

• transcripts/<filename>_summary.md — structured summary with key decisions, action items, and timeline
• transcripts/<filename>_transcription.md — full speaker-labeled transcript with timestamps

Architecture

The plugin has two parts: the skill (runs inside Claude Code) and the MCP server (runs on a host you choose).

┌─────────────────────────────────────────────────────────┐
│  Your machine (Claude Code)                             │
│                                                         │
│  /transcribe skill                                      │
│    ├── Validates input file                             │
│    ├── Extracts audio from video (ffmpeg)               │
│    ├── Estimates cost, asks for approval                │
│    ├── Transfers audio to MCP server host (SCP or cp)   │
│    ├── Calls transcribe_audio() on the server           │
│    └── Structures results into markdown                 │
│                                                         │
│  .transcribe-config.json  ← points to the server       │
└──────────────────────┬──────────────────────────────────┘
                       │ SCP + SSH (or local if same machine)
┌──────────────────────▼──────────────────────────────────┐
│  MCP Server Host (local or remote)                      │
│                                                         │
│  <install_path>/          e.g. /opt/mcp-servers/        │
│  └── transcribe-mcp/          transcribe-mcp/           │
│      ├── server.py        ← FastMCP server wrapping     │
│      │                       AssemblyAI API             │
│      ├── .venv/           ← Python virtualenv with      │
│      │                       assemblyai, python-dotenv, │
│      │                       mcp packages               │
│      ├── .env             ← ASSEMBLYAI_API_KEY=...      │
│      │                       (chmod 600)                │
│      └── requirements.txt                               │
└─────────────────────────────────────────────────────────┘

What gets installed where