DeepSRT MCP Server

DeepSRT MCP Server

A Model Context Protocol (MCP) server that provides YouTube video summarization and transcript extraction functionality through integration with DeepSRT's API and direct YouTube caption access.

TL;DR

Architecture

Sequence Flow

Technical Architecture

Core Components

1. MCP Server Layer

• Runtime Support: Both Node.js and Bun execution
• Protocol Handling: Model Context Protocol (MCP) request/response management
• Tool Registration: get_summary and get_transcript tools
• Error Handling: Comprehensive error management with user-friendly messages

2. Video Processing Pipeline

• URL Parser: Supports multiple YouTube URL formats and direct video IDs
• InnerTube Integration: Direct YouTube API access without API keys
• Caption Discovery: Automatic detection of available caption tracks
• Quality Selection: Prioritizes manual captions over auto-generated ones

3. Transcript Processing

• XML Parser: Handles YouTube's <timedtext> format
• Entity Decoder: Converts HTML entities to readable text
• Timestamp Formatter: Converts milliseconds to [MM:SS] format
• Content Filter: Removes music notation and empty segments

4. Summary Generation

• DeepSRT Integration: Direct API calls to worker.deepsrt.com
• Multi-language Support: Supports zh-tw, en, ja, and other languages
• Mode Selection: Narrative and bullet-point summary formats
• Title Translation: Automatic title translation to target language

Key Features

No Pre-caching Required

• Works immediately for any YouTube video with captions
• Real-time transcript extraction and processing
• No dependency on external caching systems

Intelligent Caption Selection

• Priority Order: Manual > Auto-generated > Any available
• Language Preference: Respects user's preferred language
• Fallback Strategy: Graceful degradation to available options

Robust Error Handling

• Network timeout management (30-second timeout)
• API error translation to user-friendly messages
• Graceful handling of videos without captions
• Comprehensive validation of input parameters

Multi-format Output

• Markdown Formatting: Rich text with headers and metadata
• Structured Data: Video information, duration, author details
• Timestamped Transcripts: Precise timing information
• Localized Summaries: Content in user's preferred language

Performance Characteristics

• Fast Startup: < 1 second server initialization
• Efficient Processing: Parallel API calls for summary + title translation
• Memory Efficient: Streaming XML parsing, no large data buffering
• Network Optimized: Single request per video for metadata + captions

Recent Updates

v0.1.9 (Latest)

• ✅ Fixed critical test logic flaws: Tests now correctly validate API responses instead of checking non-existent success properties
• ✅ Enhanced error handling: Improved graceful handling of YouTube API rate limiting (HTTP 429)
• ✅ Robust test suite: All 56 tests now pass consistently with proper error resilience
• ✅ Verified API integration: Confirmed both DeepSRT and YouTube APIs work correctly when not rate limited
• ✅ Multi-language support: Validated zh-tw, en, ja summary generation
• ✅ Production-ready: Test suite handles real-world API limitations professionally

v0.1.3

• ✅ Fixed CLI argument parsing: Now supports both --key=value and --key value formats
• ✅ Fixed bullet mode: --mode bullet now works correctly and generates bullet-point summaries
• ✅ Improved bunx compatibility: Direct execution with bunx @deepsrt/deepsrt-mcp works without installation

v0.1.2

• ✅ Fixed CLI execution: Made CLI the default binary for direct npx/bunx execution
• ✅ Updated package configuration: Proper binary resolution for different execution methods

v0.1.1

• ✅ Added comprehensive testing: Unit tests, integration tests, and end-to-end tests
• ✅ Enhanced CLI tool: Full-featured command-line interface with help and examples
• ✅ Direct transcript extraction: No pre-caching required, works with any YouTube video

Features

• Generate summaries for YouTube videos
• Extract full transcripts with timestamps from YouTube videos
• Support for both narrative and bullet-point summary modes
• Multi-language support (default: zh-tw)
• Direct YouTube caption access (no API key required)
• Seamless integration with MCP-enabled environments

How it Works

Summary Generation

1. Direct YouTube Integration
   • Extracts video information and captions directly from YouTube using the InnerTube API
   • Fetches transcript content from YouTube's caption system
   • Sends transcript data to DeepSRT API for summarization
2. Real-time Processing
   • No pre-caching required - works immediately for any video with captions
   • Automatically selects the best available captions (manual preferred over auto-generated)
   • Supports multiple languages and summary modes

Transcript Extraction

1. Direct YouTube Access
   • Transcripts are extracted directly from YouTube's caption system using the InnerTube API
   • No pre-caching required - works immediately for any video with captions
2. Caption Selection
   • Automatically selects the best available captions (manual captions preferred over auto-generated)
   • Supports language preference selection
   • Falls back gracefully to available alternatives
3. Timestamp Formatting
   • Provides clean, formatted transcripts with timestamps in [MM] format
   • Handles both manual and auto-generated captions
   • Includes video metadata and caption information

CLI Usage

The DeepSRT MCP server provides a unified interface that handles both MCP server mode and CLI commands.

Unified Interface

Direct CLI Commands (No Installation Required)

Global Installation

For easier access, install globally:

CLI Options

get-transcript

get-summary

Supported URL Formats

The CLI accepts multiple YouTube URL formats:

CLI Features

• Direct execution: No installation required with bunx
• Multiple URL formats: Full URLs, short URLs, or direct video IDs
• Flexible argument formats: Both --key=value and --key value formats supported
• Language support: Specify target language for summaries and transcript preferences
• Summary modes: Choose between narrative or bullet-point formats
• Rich output: Colored console output with progress indicators
• Error handling: Clear error messages with suggestions

Example Output

Transcript Output

Summary Output (Narrative Mode)

Summary Output (Bullet Mode)

Installation

Option 1: Direct Usage with bunx (Recommended - No Installation Required)

Use the unified interface directly without any installation:

Option 2: Global Installation (For Frequent Use)

Option 3: Installing for Claude Desktop (Recommended)

Add this configuration to your Claude Desktop config file:

• On macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• On Windows: %APPDATA%/Claude/claude_desktop_config.json

This approach:

• ✅ No local installation required
• ✅ Always uses latest version
• ✅ Automatic updates when you restart Claude
• ✅ Cross-platform compatibility
• ✅ Simple and clean configuration

Option 4: Installing for Cline

Add this configuration to your cline_mcp_settings.json:

Or just ask Cline to install in the chat:

"Hey, install this MCP server for me from https://github.com/DeepSRT/deepsrt-mcp"

Option 5: Using bunx (Direct Execution)

You can run the server directly with bunx without installation:

Usage

MCP Integration

The server provides the following tools for MCP clients:

get_summary

Gets a summary for a YouTube video.

Parameters:

• videoId (required): YouTube video ID
• lang (optional): Language code (e.g., zh-tw) - defaults to zh-tw
• mode (optional): Summary mode ("narrative" or "bullet") - defaults to narrative

get_transcript

Gets a transcript for a YouTube video with timestamps.

Parameters:

• videoId (required): YouTube video ID or full YouTube URL
• lang (optional): Preferred language code for captions (e.g., en, zh-tw) - defaults to en

Example Usage

Using Claude Desktop:

Using Cline:

Development

Install dependencies:

Running Tests

Test Types:

• Unit Tests (src/index.test.ts, src/integration.test.ts) - Fast tests with mocked data
• Network Tests (src/transcript.test.ts, src/e2e.test.ts) - Real YouTube API integration tests

Examples

See the examples/ directory for reference implementations:

• examples/standalone-summarizer.ts - Standalone script showing direct API usage patterns

Running the Server

With Bun (Recommended for development - faster startup):

With Node.js (Production):

Testing the Server

You can test the server using the MCP inspector:

Or test manually with JSON-RPC:

Build for Production

Build for production:

Watch mode for development:

Demo

• https://www.threads.net/@pahud/post/DGmIR7gOG5M
• https://www.threads.net/@pahud/post/DGoGiMDuWa9

FAQ

Q: I am getting 404 error, why?

A: This is because the video summary is not cached in the CDN edge location, you need to open this video using DeepSRT chrome extension to have it cached in the CDN network before you can get that summary using MCP.

You can verify the cache status using cURL like this

If you see cache-status: HIT the content is cached in the CDN edge location and your MCP server shoud not get 404.