augent（augent 音频）

Augent — Audio & Video Intelligence for AI Agents

Augent is an MCP server that gives your agent 22 tools for audio and video intelligence. Download from 1000+ sites via yt-dlp and aria2c, transcribe in 99 languages via faster-whisper, search by keyword or meaning via sentence-transformers, take notes, identify speakers via pyannote-audio, detect chapters, separate audio via Demucs v4, export clips, extract visual frames, record X/Twitter Spaces (requires user-configured auth token in ~/.augent/auth.json), and generate speech via Kokoro TTS. All processing runs locally. Downloads are saved to ~/Downloads/, notes and clips to ~/Desktop/, transcription memory to ~/.augent/memory/.

Config

{
  "mcpServers": {
    "augent": {
      "command": "augent-mcp"
    }
  }
}

If augent-mcp is not in PATH, use python3 -m augent.mcp as the command instead.

Install

Install via the ClawHub install button above, or use uv tool install augent for the base package or uv tool install "augent[all]" for all features. FFmpeg is required for audio processing.

Tools

Augent exposes 22 MCP tools:

Core

Analysis

Utilities

Usage Examples

Take notes from a video

"Take notes from https://youtube.com/watch?v=xxx"

The agent calls take_notes which downloads, transcribes, and returns formatted notes. One tool call does everything.

Search a podcast for topics

"Search this podcast for every mention of AI regulation" — provide the file path or URL.

The agent uses search_audio for exact keyword matches, or deep_search for semantic matches (finds relevant discussion even without exact words).

Transcribe and identify speakers

"Transcribe this meeting recording and tell me who said what"

The agent calls transcribe_audio then identify_speakers to label each segment by speaker.

Search across all transcriptions

"Search everything I've ever transcribed for mentions of funding"

The agent uses search_memory to search across all stored transcriptions without needing a file path.

Export a clip

"Clip the part where they talk about pricing"

The agent uses search_audio or deep_search to find the moment, then clip_export to extract just that segment.

Separate vocals from noisy audio

"This recording has music in the background, clean it up and transcribe"

The agent calls separate_audio to isolate vocals, then transcribe_audio on the clean vocals track.

Generate speech from text

"Read these notes aloud"

The agent calls text_to_speech to generate an MP3 with natural speech. Supports multiple voices and languages.

Note Styles

When using take_notes, the style parameter controls formatting:

Model Sizes

tiny is the default and handles nearly everything. Only use larger models for heavy accents, poor audio quality, or maximum accuracy needs.

File Paths

Augent reads and writes to these locations on your machine:

If Obsidian is installed, visual frames are saved to the Obsidian vault's External Files/visual/ directory. The vault path is auto-detected from Obsidian's config.

Network Access

Network access is used for two purposes only:

1. Downloading media from user-provided URLs via yt-dlp and aria2c
2. Downloading ML models on first use (Whisper, sentence-transformers, pyannote, Demucs, Kokoro) from Hugging Face

No telemetry. No background network activity. No data is uploaded.

ML Dependencies

The augent[all] install includes these local ML components:

All models run locally. None require API keys or cloud services.

Requirements

• Python 3.10+
• FFmpeg (audio processing)
• yt-dlp + aria2c (for audio downloads)

Links

• GitHub
• Documentation