YouTube MCP Server

Downloads video subtitles from YouTube URLs by spawning yt-dlp as a subprocess via spawn-rx, capturing VTT-formatted subtitle streams, and returning raw subtitle data to the MCP server. The implementation uses reactive streams to manage subprocess lifecycle and handle streaming output from the external command-line tool, avoiding direct HTTP requests to YouTube and instead delegating to yt-dlp's robust video metadata and subtitle retrieval logic.

Parses WebVTT (VTT) subtitle files returned by yt-dlp to extract clean, readable transcript text by removing timing metadata, cue identifiers, and formatting markup. The implementation processes line-by-line VTT content, filters out timestamp blocks (HH:MM:SS.mmm --> HH:MM:SS.mmm), and concatenates subtitle text into a continuous transcript suitable for LLM consumption, preserving speaker labels and paragraph breaks where present.

Registers YouTube subtitle extraction as a callable tool within the Model Context Protocol by defining a tool schema (name, description, input parameters) and implementing a request handler that routes incoming MCP tool_call requests to the appropriate subtitle extraction and processing logic. The implementation uses the MCP Server class to expose a single tool endpoint that Claude can invoke by name, with parameter validation and error handling integrated into the MCP request/response cycle.

Establishes a bidirectional communication channel between the mcp-youtube server and Claude.ai using the Model Context Protocol's StdioServerTransport, which reads JSON-RPC requests from stdin and writes responses to stdout. The implementation initializes the transport layer at server startup, handles the MCP handshake protocol, and maintains an event loop that processes incoming requests and dispatches responses, enabling Claude to invoke tools and receive results without explicit network configuration.

Validates incoming YouTube URLs and extracts video identifiers before passing them to yt-dlp, ensuring that only valid YouTube URLs are processed and preventing malformed or non-YouTube URLs from being passed to the subtitle extraction pipeline. The implementation likely uses regex or URL parsing to identify YouTube URL patterns (youtube.com, youtu.be, etc.) and extract the video ID, with error handling that returns meaningful error messages if validation fails.

Delegates to yt-dlp's built-in subtitle language selection and fallback logic, which automatically chooses the best available subtitle track based on user preferences, video metadata, and available caption languages. The implementation passes language preferences (if specified) to yt-dlp via command-line arguments, allowing yt-dlp to negotiate which subtitle track to download, with automatic fallback to English or auto-generated captions if the requested language is unavailable.

Catches and handles errors from yt-dlp subprocess execution, including missing binary, network failures, invalid URLs, and permission errors, returning meaningful error messages to Claude via the MCP response. The implementation wraps subprocess invocation in try-catch blocks and maps yt-dlp exit codes and stderr output to user-friendly error messages, though no explicit retry logic or exponential backoff is implemented.

Likely implements optional caching of downloaded transcripts to avoid re-downloading the same video's subtitles multiple times within a session, reducing latency and yt-dlp subprocess overhead for repeated requests. The implementation may use an in-memory cache keyed by video URL or video ID, with optional persistence to disk or external cache store, though the DeepWiki analysis does not explicitly confirm this capability.