Transcription MCP
Local transcription at 76x realtime using Apple's Neural Engine. Zero API costs. YouTube-ready captions.
Transcribe a 3-hour video in under 3 minutes. Everything runs on-device — your audio never leaves your Mac.
What It Does
- Transcribes audio and video files using Apple SpeechAnalyzer on the Neural Engine
- Generates
.srtsubtitle files ready for YouTube caption upload - Writes
.txttranscripts alongside the source file - Stores transcripts in a searchable SQLite archive with full-text search
- Extracts audio slices so Claude can listen to and verify unclear words
- Works with Claude Code, OpenClaw, and any MCP-compatible client
Install```bash
pip install transcription-mcp
Or with uv:```bash
uv pip install transcription-mcp
Quick Start
1. Add to Claude Code
Add to your ~/.claude.json:```json
{
"mcpServers": {
"transcription": {
"command": "transcription-mcp",
"args": []
}
}
}
### 2. Transcribe
In Claude Code, just ask:
> "Transcribe /path/to/my-video.mp4"
Claude will transcribe the file, clean up the output, and present the formatted transcript. An `.srt` file for YouTube captions is written automatically.
### 3. Search Past Transcripts
> "Search my transcripts for 'neural engine'"
## Features
| Feature | Details |
|---------|---------|
| **Speed** | ~76x realtime on Apple Neural Engine (M1/M2/M3/M4) |
| **Privacy** | 100% on-device — audio never leaves your Mac |
| **Captions** | YouTube-ready `.srt` files with word-level timestamps |
| **Search** | Full-text search across all past transcripts (SQLite + FTS5) |
| **Formats** | Any audio/video format supported by ffmpeg |
| **Languages** | All languages supported by Apple SpeechAnalyzer |
| **Cost** | Free forever — no API keys, no subscriptions, no cloud |
## Requirements
- **macOS 26.0** (Tahoe) or later
- **Apple Silicon** (M1/M2/M3/M4) — uses the Neural Engine
- **ffmpeg** — for audio extraction from video files```bash
brew install ffmpeg
```
- **Python 3.11+**
## How It Works```
Audio/Video File
|
v
[ffmpeg] ──> 16kHz mono WAV
|
v
[SpeechCLI] ──> Apple SpeechAnalyzer (Neural Engine)
|
v
JSON (word-level timestamps + segments)
|
v
[MCP Server] ──> .txt transcript
──> .srt subtitles
──> SQLite archive (searchable)
The Swift CLI (SpeechCLI) wraps Apple's SpeechAnalyzer framework, which runs transcription on the Neural Engine — dedicated hardware for ML inference. This is why it's so fast: it's not using your CPU or GPU.
MCP Tools Reference
transcribe
Transcribe an audio or video file. Produces .txt and .srt files next to the source.
| Parameter | Type | Default | Description |
|---|---|---|---|
file_path |
string | required | Absolute path to the audio/video file |
language |
string | "en" |
Language code (e.g., "en", "es", "ja") |
get_transcript
Retrieve a stored transcript by ID with optional time filtering.
| Parameter | Type | Default | Description |
|---|---|---|---|
transcript_id |
string | required | The transcript ID |
format |
string | "segments" |
"segments" or "words" |
start_time |
float | null | Filter to words after this time (seconds) |
end_time |
float | null | Filter to words before this time (seconds) |
get_audio_slice
Extract a short audio clip for word verification.
| Parameter | Type | Default | Description |
|---|---|---|---|
transcript_id |
string | required | The transcript ID |
start_time |
float | required | Start time in seconds |
end_time |
float | required | End time in seconds |
search_transcripts
Full-text search across all stored transcripts. Supports FTS5 syntax (AND, OR, NOT, phrases).
list_transcripts
List all stored transcripts with metadata.
FAQ
Q: Does this work on Intel Macs?
A: No. Apple SpeechAnalyzer requires Apple Silicon (M1 or later) and macOS 26.
Q: What audio/video formats are supported?
A: Anything ffmpeg can decode — MP4, MOV, MKV, MP3, M4A, WAV, FLAC, and many more.
Q: How accurate is the transcription?
A: Apple SpeechAnalyzer uses on-device neural models. Accuracy is comparable to cloud services for clear speech. Claude reviews and cleans up the output.
Q: Where are transcripts stored?
A: In ~/Library/Application Support/transcription-mcp/transcripts.db. This is a standard SQLite database you can query directly.
Q: Can I use this without Claude Code?
A: Yes — it works with any MCP client. You can also run transcription-mcp directly as a stdio MCP server.
Q: Is my audio data sent anywhere?
A: No. Everything runs locally on your Mac. No network calls, no cloud APIs, no telemetry.
License
MIT
