mcp-research-pipeline
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@mcp-research-pipeline search YouTube for 'machine learning basics' and create a NotebookLM notebook"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
mcp-research-pipeline
MCP server that unifies YouTube transcripts, YouTube search, and Google NotebookLM into a research pipeline for Claude Desktop and any MCP client.
What It Does
Extract YouTube transcripts (free, no API key needed)
Search YouTube for videos, channels, and playlists (via TranscriptAPI.com)
Create NotebookLM notebooks, add sources, ask questions, and generate deliverables (podcasts, quizzes, reports, etc.)
One-shot research pipeline: search → create notebook → add sources → ask — in a single tool call
NotebookLM acts as a free RAG system — Google pays for the analysis tokens. This MCP server lets Claude Desktop interact with it programmatically.
Related MCP server: claude-youtube-transcript
Quick Start
Prerequisites
Python 3.10+
uv (recommended) or pip
Step 1: Clone and install
git clone https://github.com/rubayatkhan/mcp-research-pipeline.git
cd mcp-research-pipeline
uv syncStep 2: Install Playwright browser
NotebookLM requires a Chromium browser for authentication. This is a one-time setup:
uv run python -m playwright install chromiumNote:
playwrightis not a standalone CLI command — it's bundled inside the project's virtual environment. Always run it withuv run python -m playwright, not justplaywright.
Step 3: Authenticate with NotebookLM (optional)
uv run notebooklm loginThis opens a browser window for Google sign-in. Your credentials are saved at ~/.notebooklm/storage_state.json and persist across server restarts.
Skip this step if you only want YouTube transcript/search tools. NotebookLM tools will return a helpful error message telling you to authenticate.
Step 4: Get a TranscriptAPI key (optional)
Sign up at transcriptapi.com to get an API key. You get 100 free credits.
Skip this step if you only need
get_transcript(which is free and keyless). The search, channel, and playlist tools require this key.
Step 5: Configure Claude Desktop
Add to your Claude Desktop config:
macOS:
~/Library/Application Support/Claude/claude_desktop_config.jsonWindows:
%APPDATA%\Claude\claude_desktop_config.json
Option A — Direct venv script (recommended, avoids path issues):
{
"mcpServers": {
"research-pipeline": {
"command": "/absolute/path/to/mcp-research-pipeline/.venv/bin/mcp-research-pipeline",
"env": {
"TRANSCRIPT_API_KEY": "your-key-here"
}
}
}
}Replace /absolute/path/to/mcp-research-pipeline with your actual clone location. Leave TRANSCRIPT_API_KEY empty or omit the env block if you don't have a key yet.
Option B — Using uv run (only if your path has no spaces):
{
"mcpServers": {
"research-pipeline": {
"command": "uv",
"args": [
"run",
"--directory", "/absolute/path/to/mcp-research-pipeline",
"python", "-m", "mcp_research_pipeline"
],
"env": {
"TRANSCRIPT_API_KEY": "your-key-here"
}
}
}
}Warning: Option B fails if your path contains spaces (e.g., iCloud Drive, OneDrive, Google Drive). Use Option A instead.
Step 6: Restart Claude Desktop
Fully quit Claude Desktop (Cmd+Q / Ctrl+Q) and reopen it. The research-pipeline server should appear in your MCP tools.
Troubleshooting
"Server disconnected" in Claude Desktop
Check the server log at ~/Library/Logs/Claude/mcp-server-research-pipeline.log (macOS). Common causes:
Error | Fix |
| Your path has spaces. Switch to Option A (direct venv script). |
| Run |
| Don't use |
Server starts then immediately disconnects | NotebookLM auth may have expired. Run |
"NotebookLM is not connected"
Run uv run notebooklm login in the project directory. This opens a browser for Google authentication.
"TRANSCRIPT_API_KEY" errors
The get_transcript tool works without any API key. Only search_youtube, get_channel_latest, get_channel_videos, and get_playlist_videos need a TranscriptAPI.com key.
Paths with spaces (iCloud, OneDrive, Google Drive)
If your project lives in a path with spaces (like ~/Library/Mobile Documents/com~apple~CloudDocs/), the uv run --directory approach will fail. Two options:
Use Option A (direct venv script path) — this always works.
Create a symlink to a path without spaces:
ln -sf "/path/with spaces/mcp-research-pipeline" ~/mcp-research-pipelineThen point Claude Desktop at
~/mcp-research-pipeline/.venv/bin/mcp-research-pipeline.
Tools (15 total)
YouTube (5 tools)
Tool | Cost | Description |
| Free | Fetch transcript from a YouTube URL or video ID |
| 1 credit | Search YouTube for videos or channels |
| Free | Get 15 most recent videos from a channel |
| 1 credit/page | Paginated list of all channel videos |
| 1 credit/page | Paginated list of playlist videos |
NotebookLM — Notebooks (5 tools)
Tool | Description |
| Create a new NotebookLM notebook |
| List all notebooks |
| Add a URL, YouTube video, or text to a notebook |
| List sources in a notebook |
| Ask a question against notebook sources (RAG) |
NotebookLM — Artifacts (4 tools)
Tool | Description |
| Generate audio, video, quiz, flashcards, report, mind_map, infographic, slide_deck, or data_table |
| List all artifacts in a notebook |
| Poll generation status |
| Download a completed artifact |
Pipeline (1 tool)
Tool | Description |
| End-to-end: search YouTube → create notebook → add sources → ask question |
Environment Variables
Variable | Required | Description |
| No | TranscriptAPI.com API key (enables YouTube search tools) |
| No | Custom path to NotebookLM auth (default: |
Development
# Install with dev dependencies
uv sync --extra dev
# Run tests
uv run pytest
# Lint
uv run ruff check src/ tests/
# Run server locally (stdio mode)
uv run python -m mcp_research_pipelineArchitecture
The server uses three design patterns:
Facade Pattern: 15 MCP tools presenting a unified interface over three different APIs
Adapter Pattern:
clients/layer wraps each third-party library behind a common async interfaceLifespan Management: FastMCP lifespan hook creates expensive clients once at startup, tears them down on shutdown
server.py (FastMCP + lifespan)
├── clients/
│ ├── youtube_transcript.py → youtube-transcript-api (sync→async)
│ ├── transcript_api.py → TranscriptAPI.com REST (httpx)
│ └── notebooklm.py → notebooklm-py (async)
├── tools/
│ ├── youtube.py (5 tools)
│ ├── notebook.py (5 tools)
│ ├── artifacts.py (4 tools)
│ └── pipeline.py (1 tool)
└── utils/
├── youtube_url.py (URL parsing)
└── errors.py (error translation)License
MIT
Available Tools
15 toolsadd_sourceB
Add a source to a NotebookLM notebook.
Inputs: value (required), notebook_id (required), source_type (required), wait. Returns: result
ask_notebookA
Ask a question against the sources in a NotebookLM notebook.
NotebookLM uses your sources as context (RAG) and provides referenced answers. This is powered by Google's infrastructure.
Inputs: question (required), notebook_id (required), source_ids. Returns: result
check_artifact_statusB
Check the generation status of a NotebookLM artifact.
Inputs: task_id (required), notebook_id (required). Returns: result
create_notebookB
Create a new NotebookLM notebook.
Inputs: name (required). Returns: result
download_artifactA
Download a completed NotebookLM artifact to a local file.
Inputs: notebook_id (required), output_path (required), artifact_type (required), artifact_id. Returns: result
generate_artifactA
Generate a NotebookLM artifact from notebook sources.
Supported artifact types: audio, video, quiz, flashcards, report, mind_map, infographic, slide_deck, data_table.
Inputs: notebook_id (required), artifact_type (required), wait, source_ids, +1 more. Returns: result
get_channel_latestB
Get the 15 most recent videos from a YouTube channel. FREE — no credits.
Inputs: channel (required). Returns: result
get_channel_videosA
Paginated list of all videos from a YouTube channel. Costs 1 credit/page.
Inputs: channel, continuation. Returns: result
get_playlist_videosB
Paginated list of videos in a YouTube playlist. Costs 1 credit/page.
Inputs: playlist, continuation. Returns: result
get_transcriptB
Fetch the transcript of a YouTube video.
Inputs: video (required), languages, preserve_formatting. Returns: result
list_artifactsA
List all artifacts in a NotebookLM notebook.
Inputs: notebook_id (required). Returns: result
list_notebooksA
List all NotebookLM notebooks.
Returns: JSON array of notebooks with notebook_id, name, created_at, source_count.
Inputs: no named parameters. Returns: result
list_sourcesA
List all sources in a NotebookLM notebook.
Inputs: notebook_id (required). Returns: result
research_topicA
End-to-end research pipeline: search YouTube, create a NotebookLM notebook, add video sources, and optionally ask a question.
This tool orchestrates the full workflow:
Search YouTube for videos matching the query
Create a new NotebookLM notebook
Add each video as a YouTube source
Optionally ask a question against all sources
Requires both TRANSCRIPT_API_KEY (for search) and NotebookLM login.
Inputs: query (required), question, max_videos, notebook_name. Returns: result
search_youtubeA
Search YouTube for videos or channels. Costs 1 TranscriptAPI credit.
Inputs: query (required), limit, result_type. Returns: result
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 15 tools
Tools are mostly distinct by resource and action, with clear separation between YouTube retrieval and NotebookLM management. The main ambiguity is between get_channel_latest and get_channel_videos, though descriptions clarify the convenience vs. paginated distinction. research_topic overlaps with the individual steps but is clearly positioned as an orchestrator.
All tool names use snake_case and generally follow a verb_noun pattern, which keeps the set predictable. Minor inconsistencies like get_channel_latest omitting a noun and mixing list_/get_ for similar operations prevent a perfect score.
At 15 tools, the server sits at the upper bound of a well-scoped set, but the count is justified by covering two integrated domains: YouTube retrieval and NotebookLM management. research_topic and get_channel_latest add some redundancy but still serve useful convenience purposes.
The core research workflow is covered end-to-end: search YouTube, create a notebook, add sources, ask questions, and generate/download artifacts. Missing delete/update operations for notebooks, sources, and artifacts are notable gaps but not critical for the stated pipeline purpose.
Maintenance
Related MCP Connectors
Your YouTube library, in the AI tools you already use: Claude, ChatGPT, Cursor, any MCP client.
Your YouTube library in Claude, ChatGPT and Cursor: transcripts, breakdowns, summaries, search.
Your YouTube library in Claude, ChatGPT and Cursor: transcripts, breakdowns, summaries, search.
YouTube transcripts, search, channel/playlist listings and upload tracking for AI agents.
Related MCP Servers
- FlicenseBqualityCmaintenanceExtracts YouTube transcripts and performs AI-powered video analysis for Claude Desktop, enabling transcript retrieval, quality analysis, and smart resource management.8-
- AlicenseNot gradedqualityCmaintenanceEnables Claude Desktop to fetch full YouTube video transcripts from any URL, so users can summarize, translate, or cite video content without API keys or downloads.MIT
- FlicenseNot gradedqualityAmaintenanceEnables MCP clients such as Claude and Copilot to turn YouTube videos into a persistent, compounding wiki by extracting knowledge from transcripts and scene-change vision analysis. Lets clients run agentic search over that compiled knowledge, answering questions with source citations and awareness of what the wiki does not yet cover.-
- FlicenseNot gradedqualityAmaintenanceTurns YouTube videos into a persistent, cross-referenced wiki by extracting transcripts and describing scene-change frames with vision models. Enables MCP clients like Claude and Copilot to search and reason over this compounding knowledge base without needing an API key.-