mcp-video

Video intelligence MCP server for AI-powered video analysis. Extract keyframes, transcripts, and visual content from any video URL — right inside Claude Code or any MCP client.

<p align="center"> <img src="assets/architecture.svg" alt="mcp-video architecture" width="860"/> </p>

What it does

Point it at any video (YouTube, Vimeo, Twitter/X, Instagram, and 1000+ sites) and get:

• Keyframe extraction — scene-change detection picks the most meaningful frames
• Transcript extraction — auto-captions or manual subtitles with timestamps
• Full analysis — frames + transcript together for comprehensive video understanding
• Timestamp snapshots — grab a frame at any specific moment
• Video-to-strategy — extract trading strategies from YouTube/Instagram tutorials and generate runnable Python code (backtrader, zipline, pandas-ta, freqtrade)

Works with Claude's vision capabilities to understand what's actually shown in the video, not just what's said.

Trading Strategy Extraction

Point it at any trading tutorial video and get a complete, runnable strategy:

Convert this trading strategy to code: https://youtube.com/watch?v=...

Extract the RSI strategy from this video and generate backtrader code: https://youtube.com/watch?v=...

It extracts chart patterns, indicators, entry/exit rules from the video frames and transcript, then generates code for your preferred framework.

Install

# Requires Python 3.10+ and ffmpeg
pip install mcp-video

# Or with uv
uv pip install mcp-video

System dependencies

# macOS
brew install ffmpeg yt-dlp

# Ubuntu/Debian
sudo apt install ffmpeg
pip install yt-dlp

# Windows
choco install ffmpeg yt-dlp

Usage with Claude Code

Add to your Claude Code MCP config (~/.claude/claude_code_config.json):

{
  "mcpServers": {
    "video": {
      "command": "mcp-video",
      "args": []
    }
  }
}

Then in Claude Code, just ask:

Analyze this video: https://youtube.com/watch?v=...

What code is shown in this tutorial? https://youtube.com/watch?v=...

Summarize the key points from this talk: https://youtube.com/watch?v=...

Usage with any MCP client

# Run as stdio server (default)
mcp-video

# Run as HTTP server
python -m mcp_video.server --transport streamable-http

Tools

analyze_video

Full video analysis — extracts keyframes and transcript, returns images + text.

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | url | string | required | Video URL | | max_frames | int | 8 | Max keyframes to extract | | include_transcript | bool | true | Include captions |

extract_video_frames

Extract frames only, with keyframe detection or uniform spacing.

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | url | string | required | Video URL | | max_frames | int | 10 | Number of frames | | mode | string | "keyframe" | "keyframe" or "uniform" |

get_video_transcript

Get timestamped transcript/captions.

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | url | string | required | Video URL |

get_frame_at_time

Extract a single frame at a specific timestamp.

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | url | string | required | Video URL | | timestamp | float | required | Time in seconds |

video_to_strategy

Extract trading strategies from tutorial videos and generate runnable code.

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | url | string | required | Video URL with trading strategy | | framework | string | "backtrader" | Target: backtrader, zipline, pandas-ta, freqtrade | | max_frames | int | 15 | More frames = more chart detail captured |

How it works

1. Download — Uses yt-dlp to download video (capped at 720p to save bandwidth)
2. Keyframe detection — OpenCV histogram comparison detects scene changes
3. Frame encoding — Frames resized to 512px width, encoded as JPEG base64
4. Transcript — Extracts YouTube auto-captions or manual subtitles
5. MCP response — Returns ImageContent + TextContent for the LLM to analyze

All processing happens locally. No data sent to third-party services.

Development

git clone https://github.com/ipythonist/mcp-video.git
cd mcp-video
uv venv && uv pip install -e ".[dev]"

# Run tests
pytest

# Lint
ruff check src/

License

MIT