OcularAudio MCP

An asynchronous Model Context Protocol (MCP) server that gives AI models "eyes and ears" to process web videos. It extracts transcripts and captures screenshots from YouTube and other video platforms.

Features

• Hybrid transcript extraction: Fetches YouTube captions instantly, falls back to local Whisper ASR
• On-demand video screenshots: Captures frames at any timestamp without downloading the full video
• OCR on screenshots: Extract visible text from frames using Tesseract (optional, --ocr flag)
• Cookie authentication: Supports age-restricted and private videos via cookies.txt
• Local caching: Processed videos are cached for instant subsequent lookups
• Async architecture: Non-blocking design keeps MCP clients responsive
• Flexible output: Clipboard, stdout, file, or JSON — your choice

Benchmark

See BENCHMARK.md for performance benchmarks and a deep competitive analysis against all major video transcript, screenshot, and OCR tools in the MCP and CLI ecosystem.

Requirements

• Python 3.9+ (required for list[int] type hints)
• FFmpeg (required by yt-dlp and OpenCV)
• Node.js 18+ (only for the CLI wrapper)
• Tesseract (optional, only for --ocr flag)

Installation

1. Install system dependencies

macOS:

brew install ffmpeg python3
# Optional (for OCR):
brew install tesseract

Windows:

choco install ffmpeg python
# Optional (for OCR):
choco install tesseract

Linux:

sudo apt update && sudo apt install ffmpeg python3 python3-pip
# Optional (for OCR):
sudo apt install tesseract-ocr

2. Install Python packages

pip install -r requirements.txt

Or manually:

pip install mcp youtube-transcript-api yt-dlp opencv-python-headless faster-whisper requests pytesseract

3. Install Node.js CLI (optional)

npm install

Usage

Option A: MCP Server (Recommended)

The MCP server gives AI models direct access to video transcripts and screenshots.

Quick Install — No installation needed. Just add the config to your MCP client below.

Claude Desktop

{
  "mcpServers": {
    "ocular-audio-mcp": {
      "command": "npx",
      "args": ["ocular-audio-mcp"]
    }
  }
}

Config: ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows)

Cursor

{
  "mcpServers": {
    "ocular-audio-mcp": {
      "command": "npx",
      "args": ["ocular-audio-mcp"]
    }
  }
}

Config: .cursor/mcp.json (project) or ~/.cursor/mcp.json (global)

Claude Code

claude mcp add ocular-audio-mcp -- npx ocular-audio-mcp

Codex CLI (OpenAI)

codex mcp add ocular-audio-mcp -- npx ocular-audio-mcp

Gemini CLI

gemini mcp add ocular-audio-mcp npx ocular-audio-mcp --scope user

Windsurf

{
  "mcpServers": {
    "ocular-audio-mcp": {
      "command": "npx",
      "args": ["ocular-audio-mcp"]
    }
  }
}

Config: ~/.codeium/windsurf/mcp_config.json

Zed

{
  "context_servers": {
    "ocular-audio-mcp": {
      "command": "npx",
      "args": ["ocular-audio-mcp"]
    }
  }
}

Config: ~/.config/zed/settings.json

VS Code (GitHub Copilot)

{
  "servers": {
    "ocular-audio-mcp": {
      "type": "stdio",
      "command": "npx",
      "args": ["ocular-audio-mcp"]
    }
  }
}

Config: .vscode/mcp.json

OpenCode

{
  "mcpServers": {
    "ocular-audio-mcp": {
      "command": "npx",
      "args": ["ocular-audio-mcp"]
    }
  }
}

Config: ~/.opencode/config.json

Cline (VS Code Extension)

{
  "mcpServers": {
    "ocular-audio-mcp": {
      "command": "npx",
      "args": ["ocular-audio-mcp"]
    }
  }
}

Local Development (from source)

{
  "mcpServers": {
    "ocular-audio-mcp": {
      "command": "python",
      "args": ["/path/to/ocular_audio_mcp.py"]
    }
  }
}

Option B: CLI

npx ocular-audio "https://www.youtube.com/watch?v=dQw4w9WgXcQ"

By default, the transcript is printed to stdout and copied to your clipboard. Paste it into Claude Web, ChatGPT, or any AI chat.

CLI Options

Examples

# Basic usage — prints to stdout + copies to clipboard
npx ocular-audio "https://www.youtube.com/watch?v=dQw4w9WgXcQ"

# Stdout only — great for piping to other tools
npx ocular-audio --stdout "https://www.youtube.com/watch?v=dQw4w9WgXcQ" | head -50

# Write to a specific file
npx ocular-audio --output transcript.txt "https://www.youtube.com/watch?v=dQw4w9WgXcQ"

# Raw JSON for programmatic consumption
npx ocular-audio --json "https://www.youtube.com/watch?v=dQw4w9WgXcQ" | jq .metadata.title

# Transcript only, no screenshots
npx ocular-audio --detail overview "https://www.youtube.com/watch?v=dQw4w9WgXcQ"

# Maximum screenshots
npx ocular-audio --detail deep "https://www.youtube.com/watch?v=dQw4w9WgXcQ"

# No clipboard copy, just print to terminal
npx ocular-audio --no-clipboard "https://www.youtube.com/watch?v=dQw4w9WgXcQ"

# Screenshots with OCR — extract visible text from frames
npx ocular-audio --ocr "https://www.youtube.com/watch?v=dQw4w9WgXcQ"

# JSON output with OCR
npx ocular-audio --json --ocr "https://www.youtube.com/watch?v=dQw4w9WgXcQ"

# Force re-process (bypass cache)
npx ocular-audio --force "https://www.youtube.com/watch?v=dQw4w9WgXcQ"

# Verbose mode — see all progress details
npx ocular-audio --verbose "https://www.youtube.com/watch?v=dQw4w9WgXcQ"

# Quiet mode — minimal output
npx ocular-audio --quiet --stdout "https://www.youtube.com/watch?v=dQw4w9WgXcQ"

# Check system capabilities
npx ocular-audio --check

# List cached videos
npx ocular-audio --list-cached

# Show cache stats
npx ocular-audio --cache-info

# Clear all cached data
npx ocular-audio --clear-cache

Cookie Setup (for age-restricted/private videos)

YouTube may block transcript access for age-restricted or private videos. To fix this, export your browser cookies:

1. Install a browser extension like "Get cookies.txt LOCALLY" (Chrome/Firefox)
2. Go to youtube.com while logged in
3. Export cookies to a file named cookies.txt
4. Place the file in one of these locations:
   • ~/.cache/ocular_audio_mcp/cookies.txt
   • ~/.config/ocular_audio_mcp/cookies.txt
   • ./cookies.txt (in the project directory)

The server will automatically detect and use the cookies file.

MCP Tools

get_ocular_audio_capabilities

Returns system capabilities and dependency status. Use this to check what features are available.

Parameters: None

Returns: System info including Python version, FFmpeg, Whisper, Tesseract, OpenCV, and cookie status.

get_ocular_audio_metadata

Extracts only video metadata (title, creator, duration, views, chapters) without transcript. Much faster than getting the full transcript.

Parameters:

• url (string): Video URL

get_ocular_audio_transcript

Extracts the complete transcript, video chapters, and metadata from a video.

Parameters:

• url (string): Video URL
• use_local_whisper (boolean, default: true): Enable Whisper fallback if captions unavailable

get_ocular_audio_chapters

Extracts only video chapters with timestamps. Returns chapter titles with start times in [MM:SS] format.

Parameters:

• url (string): Video URL

get_ocular_audio_video_screenshots

Captures screenshots at specific timestamps.

Parameters:

• url (string): Video URL
• timestamps_secs (array of integers): Timestamps to capture (e.g., [45, 120, 300])
• enable_ocr (boolean, default: false): If true, run OCR on each captured frame to extract visible text

get_ocular_audio_video_context

Extracts transcript, metadata, and intelligent screenshots in one call. Automatically analyzes the transcript to find visually important moments and captures screenshots at those timestamps.

Parameters:

• url (string): Video URL
• detail_level (string, default: "auto"): Controls screenshot capture mode:
  • "auto" - Adapts to video length and content importance
  • "overview" - Transcript and metadata only, no screenshots (fastest)
  • "balanced" - Screenshots only at visually important moments (strong signals)
  • "deep" - Screenshots at every visually significant moment (all signals)
• use_local_whisper (boolean, default: true): Enable Whisper fallback if captions unavailable
• enable_ocr (boolean, default: false): If true, run OCR on captured screenshots to extract visible text

list_ocular_audio_cache

Lists all cached videos with their metadata (title, uploader, duration, when cached).

Parameters: None

clear_ocular_audio_cache

Clears cached video data.

Parameters:

• video_id (string, optional): Video ID to clear specific video. If empty, clears all cache.

Cache Management