DeepSRT MCP Server by 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

{
  "mcpServers": {
    "deepsrt": {
      "type": "stdio",
      "command": "bunx",
      "args": [
        "@deepsrt/deepsrt-mcp@latest",
        "--server"
      ]
    }
  }
}

Architecture

graph TB
    subgraph "MCP Client"
        Client[Claude Desktop / Cline]
    end
    
    subgraph "DeepSRT MCP Server"
        Server[MCP Server]
        SummaryTool[get_summary]
        TranscriptTool[get_transcript]
        
        Server --> SummaryTool
        Server --> TranscriptTool
    end
    
    subgraph "External APIs"
        YouTube[YouTube InnerTube API]
        DeepSRT[DeepSRT Worker API]
        Captions[YouTube Caption API]
    end
    
    Client --> Server
    SummaryTool --> YouTube
    SummaryTool --> DeepSRT
    TranscriptTool --> YouTube
    TranscriptTool --> Captions

Sequence Flow

sequenceDiagram
    participant User
    participant MCP as MCP Client
    participant Server as DeepSRT MCP Server
    participant YouTube as YouTube InnerTube API
    participant DeepSRT as DeepSRT Worker API
    participant Captions as YouTube Caption API

    Note over User,Captions: Summary Generation Flow
    User->>MCP: Request video summary
    MCP->>Server: get_summary(videoId, lang, mode)
    
    Server->>Server: Extract video ID from URL
    Server->>YouTube: POST /youtubei/v1/player
    Note right of YouTube: Get video metadata<br/>and caption tracks
    YouTube-->>Server: Video details + caption tracks
    
    Server->>Server: Select best caption track<br/>(manual > auto-generated)
    Server->>Server: Extract transcript argument<br/>from caption URL
    
    par Summary Request
        Server->>DeepSRT: GET /transcript2?action=summarize
        Note right of DeepSRT: X-Transcript-Arg header<br/>contains caption URL params
        DeepSRT-->>Server: Generated summary
    and Title Translation
        Server->>DeepSRT: GET /transcript2?action=translate
        DeepSRT-->>Server: Translated title
    end
    
    Server->>Server: Format markdown response<br/>with metadata + summary
    Server-->>MCP: Formatted summary response
    MCP-->>User: Display summary

    Note over User,Captions: Transcript Extraction Flow
    User->>MCP: Request video transcript
    MCP->>Server: get_transcript(videoId, lang)
    
    Server->>Server: Extract video ID from URL
    Server->>YouTube: POST /youtubei/v1/player
    YouTube-->>Server: Video details + caption tracks
    
    Server->>Server: Select best caption track<br/>for preferred language
    Server->>Captions: GET caption XML from baseUrl
    Captions-->>Server: Raw XML transcript
    
    Server->>Server: Parse XML transcript<br/>- Extract timestamps<br/>- Decode HTML entities<br/>- Format text
    Server->>Server: Generate markdown response<br/>with timestamps
    Server-->>MCP: Formatted transcript
    MCP-->>User: Display transcript with timestamps

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

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
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

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
Caption Selection
- Automatically selects the best available captions (manual captions preferred over auto-generated)
- Supports language preference selection
- Falls back gracefully to available alternatives
Timestamp Formatting
- Provides clean, formatted transcripts with timestamps in [MM:SS] format
- Handles both manual and auto-generated captions
- Includes video metadata and caption information

%%{init: {'theme': 'dark', 'themeVariables': { 'primaryColor': '#2496ED', 'secondaryColor': '#38B2AC', 'tertiaryColor': '#1F2937', 'mainBkg': '#111827', 'textColor': '#E5E7EB', 'lineColor': '#4B5563', 'noteTextColor': '#E5E7EB'}}}%%
sequenceDiagram
    participant User
    participant MCP as MCP Client
    participant YouTube as YouTube API
    participant DeepSRT as DeepSRT API

    Note over User,DeepSRT: Summary Generation Flow (Direct Processing)
    User->>MCP: Request video summary
    MCP->>YouTube: Get video info & captions via InnerTube API
    YouTube-->>MCP: Return video details & caption tracks
    MCP->>YouTube: Fetch transcript XML from caption URL
    YouTube-->>MCP: Return raw transcript content
    MCP->>DeepSRT: Send transcript + metadata for summarization
    DeepSRT-->>MCP: Return generated summary
    MCP-->>User: Return formatted summary

    Note over User,DeepSRT: Transcript Extraction Flow (Direct)
    User->>MCP: Request video transcript
    MCP->>YouTube: Get video info & captions via InnerTube API
    YouTube-->>MCP: Return video details & caption tracks
    MCP->>YouTube: Fetch transcript XML from caption URL
    YouTube-->>MCP: Return raw transcript content
    MCP->>MCP: Parse & format transcript with timestamps
    MCP-->>User: Return formatted transcript with metadata

CLI Usage

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

Unified Interface

# MCP Server Mode (default - for Claude Desktop/Cline)
bunx @deepsrt/deepsrt-mcp                    # Starts MCP server on stdio
bunx @deepsrt/deepsrt-mcp --server           # Explicit server mode

# CLI Commands (direct usage)
bunx @deepsrt/deepsrt-mcp get-transcript <video-url> [options]
bunx @deepsrt/deepsrt-mcp get-summary <video-url> [options]

# Help
bunx @deepsrt/deepsrt-mcp --help

Direct CLI Commands (No Installation Required)

# Extract transcript with timestamps
bunx @deepsrt/deepsrt-mcp get-transcript https://www.youtube.com/watch?v=dQw4w9WgXcQ
bunx @deepsrt/deepsrt-mcp get-transcript dQw4w9WgXcQ --lang=en

# Generate video summary
bunx @deepsrt/deepsrt-mcp get-summary dQw4w9WgXcQ --lang=zh-tw --mode=bullet
bunx @deepsrt/deepsrt-mcp get-summary https://youtu.be/dQw4w9WgXcQ --lang=ja

Global Installation

For easier access, install globally:

# Install globally
npm install -g @deepsrt/deepsrt-mcp

# MCP Server Mode
deepsrt-mcp                                             # Starts MCP server
deepsrt-mcp --server                                    # Explicit server mode

# CLI Commands
deepsrt-mcp get-transcript https://youtu.be/dQw4w9WgXcQ --lang=en
deepsrt-mcp get-summary dQw4w9WgXcQ --lang=zh-tw --mode=narrative

CLI Options

get-transcript

bunx @deepsrt/deepsrt-mcp get-transcript <video-url> [options]

Options:
  --lang=<language>    Preferred language code for captions (default: en)
  --lang <language>    Alternative format
                      Examples: en, zh-tw, ja, es, fr

Examples:
  bunx @deepsrt/deepsrt-mcp get-transcript https://www.youtube.com/watch?v=dQw4w9WgXcQ
  bunx @deepsrt/deepsrt-mcp get-transcript dQw4w9WgXcQ --lang=zh-tw
  bunx @deepsrt/deepsrt-mcp get-transcript https://youtu.be/dQw4w9WgXcQ --lang ja

get-summary

bunx @deepsrt/deepsrt-mcp get-summary <video-url> [options]

Options:
  --lang=<language>    Target language for summary (default: zh-tw)
  --lang <language>    Alternative format
                      Examples: zh-tw, en, ja, es, fr
  --mode=<mode>        Summary format (default: narrative)
  --mode <mode>        Alternative format
                      Options: narrative, bullet

Examples:
  bunx @deepsrt/deepsrt-mcp get-summary https://www.youtube.com/watch?v=dQw4w9WgXcQ
  bunx @deepsrt/deepsrt-mcp get-summary dQw4w9WgXcQ --lang=en --mode=bullet
  bunx @deepsrt/deepsrt-mcp get-summary https://youtu.be/dQw4w9WgXcQ --lang ja --mode narrative

Supported URL Formats

The CLI accepts multiple YouTube URL formats:

# Full YouTube URLs
https://www.youtube.com/watch?v=dQw4w9WgXcQ
https://www.youtube.com/watch?v=dQw4w9WgXcQ&t=30s

# Short URLs
https://youtu.be/dQw4w9WgXcQ

# Embed URLs
https://www.youtube.com/embed/dQw4w9WgXcQ

# Direct video IDs
dQw4w9WgXcQ

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

# Rick Astley - Never Gonna Give You Up (Official Video) (4K Remaster)

**Author:** Rick Astley  
**Duration:** 3:33  
**Captions:** English (manual)

## Transcript

[00:18] ♪ We're no strangers to love ♪
[00:22] ♪ You know the rules and so do I ♪
[00:27] ♪ A full commitment's what I'm thinking of ♪
...

Summary Output (Narrative Mode)

# 瑞克·艾斯利 - 永遠不會放棄你

**Author:** Rick Astley  
**Duration:** 3:33  
**Language:** zh-tw  
**Mode:** narrative

## Summary

這是一首經典的流行歌曲，表達了對愛情的承諾和忠誠...

Summary Output (Bullet Mode)

# How Robots Are Helping Amazon Deliver on Prime Day

**Author:** Bloomberg Television  
**Duration:** 5:45  
**Language:** zh-tw  
**Mode:** bullet

## Summary

本影片主要探討亞馬遜如何運用機器人技術來提升倉儲效率...

# 機器人技術在亞馬遜倉儲的應用與效益

  - 亞馬遜已部署了第一百萬個機器人，並在Prime Day等高峰期用於滿足訂單需求 [00:00:00]
  - 機器人旨在為員工提供安全且高生產力的工作環境 [00:00:30]
  - 移動機器人可以搬運超過一千磅的貨物，並移動貨架以減少員工的行走距離 [00:00:54]
...

Installation

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

Use the unified interface directly without any installation:

# MCP Server Mode (for Claude Desktop/Cline)
bunx @deepsrt/deepsrt-mcp                    # Default: starts MCP server
bunx @deepsrt/deepsrt-mcp --server           # Explicit server mode

# CLI Commands (direct usage)
bunx @deepsrt/deepsrt-mcp get-transcript https://www.youtube.com/watch?v=dQw4w9WgXcQ
bunx @deepsrt/deepsrt-mcp get-summary dQw4w9WgXcQ --lang=zh-tw --mode=bullet

# Always uses latest version automatically
bunx @deepsrt/deepsrt-mcp@latest get-transcript dQw4w9WgXcQ --lang=en

Option 2: Global Installation (For Frequent Use)

# Install globally for easier access
npm install -g @deepsrt/deepsrt-mcp

# Then use directly with shorter command
deepsrt-mcp get-transcript https://www.youtube.com/watch?v=dQw4w9WgXcQ
deepsrt-mcp get-summary dQw4w9WgXcQ --lang zh-tw --mode bullet

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

{
  "mcpServers": {
    "deepsrt": {
      "type": "stdio",
      "command": "bunx",
      "args": [
        "@deepsrt/deepsrt-mcp@latest",
        "--server"
      ]
    }
  }
}

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:

{
  "mcpServers": {
    "deepsrt": {
      "type": "stdio", 
      "command": "bunx",
      "args": [
        "@deepsrt/deepsrt-mcp@latest",
        "--server"
      ]
    }
  }
}

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:

# Run from the project directory
bunx --bun src/index.ts

# Or use npm scripts
npm run start:bun  # Uses Bun
npm run start:node # Uses Node.js
npm run dev        # Development mode with Bun

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:

// Get video summary
const summaryResult = await mcp.use_tool("deepsrt", "get_summary", {
  videoId: "dQw4w9WgXcQ",
  lang: "zh-tw",
  mode: "narrative"
});

// Get video transcript
const transcriptResult = await mcp.use_tool("deepsrt", "get_transcript", {
  videoId: "dQw4w9WgXcQ",
  lang: "en"
});

Using Cline:

// Get video summary
const summaryResult = await mcp.use_tool("deepsrt", "get_summary", {
  videoId: "dQw4w9WgXcQ",
  lang: "zh-tw",
  mode: "bullet"
});

// Get video transcript
const transcriptResult = await mcp.use_tool("deepsrt", "get_transcript", {
  videoId: "https://www.youtube.com/watch?v=dQw4w9WgXcQ",
  lang: "en"
});

Development

Install dependencies:

npm install

Running Tests

# Run unit tests (fast, no network calls)
npm test

# Run unit tests only
npm run test:unit

# Run network tests (requires internet, may be slower)
npm run test:network

# Run all tests including network tests
npm run test:all

# Run tests in watch mode
npm run test:watch

# Run tests with CI reporter (for CI/CD)
npm run test:ci

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):

# Development mode (runs TypeScript directly)
npm run dev
# or
bun src/index.ts

# Using npm script
npm run start:bun

With Node.js (Production):

# Build first
npm run build

# Then run
npm run start:node
# or
node build/index.js

Testing the Server

You can test the server using the MCP inspector:

npm run inspector

Or test manually with JSON-RPC:

# List available tools
echo '{"jsonrpc": "2.0", "id": 1, "method": "tools/list"}' | bun src/index.ts

# Test get_transcript
echo '{"jsonrpc": "2.0", "id": 1, "method": "tools/call", "params": {"name": "get_transcript", "arguments": {"videoId": "dQw4w9WgXcQ", "lang": "en"}}}' | bun src/index.ts

Build for Production

Build for production:

npm run build

Watch mode for development:

npm run watch

Demo

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

curl -s 'https://worker.deepsrt.com/transcript' \
-i --data '{"arg":"v=VafNvIcOs5w","action":"summarize","lang":"zh-tw","mode":"narrative"}' | grep -i "^cache-status"
cache-status: HIT

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