YouTube Research MCP

Turn YouTube videos into citation-ready research for Codex, Claude, Cursor, and other MCP clients. Paste a video URL and get transcript evidence with timestamps and links that open at the exact quoted moment.

Input:  https://youtu.be/dQw4w9WgXcQ
Output: [01:05] ...evidence text...
        https://www.youtube.com/watch?v=dQw4w9WgXcQ&t=65s

No YouTube API key is required for transcript research.

Why this server

• Citation-ready research — every transcript segment includes a timestamp and source URL.
• Zero-key quick start — transcripts work immediately; add an API key only for search and analytics.
• URL-first input — accepts normal, short, Shorts, embed, live, and raw video-ID formats.
• Agent-efficient filtering — search inside transcripts and return nearby context instead of spending tokens on the entire video.
• Full YouTube intelligence — optional API mode adds comments, video/channel statistics, trends, and comparisons.
• Remote-native MCP — Streamable HTTP transport at /mcp, plus Docker and Smithery support.

Architecture and trust boundary

flowchart LR
    C[MCP client] -->|stdio or Streamable HTTP| T[Transport]
    T --> R[Tool and resource registry]
    R --> S[YouTubeService]
    S --> K[(In-memory cache)]
    S -->|no key required| YT[Caption tracks]
    S -->|optional API key| API[YouTube Data API v3]

The MCP layer validates inputs and returns structured evidence; it does not ask a model to invent summaries or citations. Transcript tools derive timestamps and canonical links from caption segments. Data API tools return an explicit key-required error when YOUTUBE_API_KEY is unavailable, and remote authentication is enforced before MCP session handling.

Focused evidence, not a transcript dump

The included live smoke test asks a focused question about a public video and compares the response with the full timestamped transcript:

This measures response characters, not model-specific tokens. Reproduce it against the default public fixture—or substitute your own video and query:

npm run test:live
LIVE_TEST_VIDEO='https://youtu.be/VIDEO_ID' LIVE_TEST_QUERY='evaluation' npm run test:live

Quick start — no API key

Requirements: Node.js 20 or newer.

The shortest local install uses stdio and needs no API key:

codex mcp add youtube-research -- npx -y @coyasong/youtube-mcp-server@latest

For Claude Desktop and other MCPB-compatible desktop clients, download the latest one-click MCP bundle and open it. The bundle vendors its runtime dependencies, starts locally over stdio, and asks for a YouTube API key only if you want the optional analytics tools.

Or install from the VS Code command line:

code --add-mcp '{"name":"youtube-research","command":"npx","args":["-y","@coyasong/youtube-mcp-server@latest"]}'

For Claude Desktop, Cursor, and other stdio clients, use command npx with arguments -y @coyasong/youtube-mcp-server@latest.

To run the Streamable HTTP transport instead:

git clone https://github.com/coyaSONG/youtube-mcp-server.git
cd youtube-mcp-server
npm install
npm run build
npm start

The server starts at http://localhost:3000/mcp in transcript-only mode. Confirm it with:

curl http://localhost:3000/health

Connect from Codex

With the HTTP server running:

codex mcp add youtube-research --url http://localhost:3000/mcp

Then ask Codex:

Use research-video to find what this video says about evaluation,
and cite the exact moments: https://www.youtube.com/watch?v=VIDEO_ID

Connect with MCP Inspector

npx @modelcontextprotocol/inspector
# Connect to http://localhost:3000/mcp

Install through Smithery

npx -y smithery@latest auth login
npx -y smithery@latest mcp add coyaSONG/youtube-mcp-server --client claude

Enable search, comments, and analytics

Set a YouTube Data API v3 key to switch from transcript-only to full mode:

export YOUTUBE_API_KEY=your_key_here
npm start

YOUTUBE_API_KEY is optional. PORT defaults to 3000.

Secure a remote deployment

Do not expose a full-mode server publicly without authentication: unauthenticated users could consume your YouTube API quota. Set a strong bearer token and restrict browser origins when deploying outside localhost:

export MCP_BEARER_TOKEN='replace-with-a-long-random-secret'
export CORS_ORIGIN='https://your-client.example'
export MAX_SESSIONS=100
export SESSION_IDLE_TIMEOUT_MS=1800000
npm start

Browser origins are denied by default. CORS_ORIGIN accepts a comma-separated allowlist of exact http:// or https:// origins; wildcard origins are rejected.

Connect Codex using an environment variable rather than writing the secret into its configuration:

export YOUTUBE_MCP_TOKEN='replace-with-a-long-random-secret'
codex mcp add youtube-research \
  --url https://your-server.example/mcp \
  --bearer-token-env-var YOUTUBE_MCP_TOKEN

/health remains public for container health checks. MCP requests return 401 when authentication is enabled and the bearer token is missing or invalid. Idle sessions are removed automatically, and MAX_SESSIONS bounds memory use.

Best first tool

research-video accepts:

• video: a YouTube URL or 11-character video ID
• language: optional caption language such as en, ko, or ja
• query: optional phrase to find inside the transcript
• contextLines: surrounding segments to retain, from 0 to 5
• matchMode: word (default) or substring
• startSeconds / endSeconds: optional time window
• offset: result offset for pagination
• maxSegments: response cap from 1 to 1,000 (default: 200)

It returns structured JSON containing the video title and channel identity, canonical video URL, full caption-track duration and segment count, matching transcript segments, human-readable citation labels, timestamps, directly navigable citation URLs, and pagination metadata. For long videos, use a query or time window first; follow nextOffset only when more evidence is needed.

Compare evidence across videos

research-videos applies one focused query to 2–5 video URLs concurrently. It returns the same structured, timestamp-linked evidence for each source while capping results per video. This is useful for comparing interviews, checking whether multiple sources support a claim, or researching a topic across a short watchlist.

{
  "videos": [
    "https://youtu.be/VIDEO_ONE",
    "https://youtu.be/VIDEO_TWO"
  ],
  "query": "evaluation",
  "maxSegmentsPerVideo": 10
}

See YouTube Research Recipes for copy-paste workflows for fact-checking claims, comparing interviews, navigating long talks, researching multilingual captions, and building citation-ready notes.

Capability modes

Captions must be available for the requested video. Age-restricted, private, region-restricted, or caption-disabled videos may not return a transcript.

If this project saves you research time, consider starring the repository so other agent builders can discover it.

See the broader AI tooling portfolio for the reliability principles shared with ralph-research and tmuxicate.

Docker

docker build -t youtube-research-mcp .
docker run --rm -p 3000:3000 youtube-research-mcp

# Full mode
docker run --rm -p 3000:3000 -e YOUTUBE_API_KEY=your_key youtube-research-mcp

Development

npm run dev             # HTTP server from TypeScript
npm test                # build and run all tests
npm run test:live       # live public-video transcript and citation smoke test
npm run test:user       # clean-room smoke test against the published npm package
npm run test:mcpb       # build, unpack, and exercise the installable MCP bundle

Maintainers can follow the release guide. Published GitHub releases run the complete npm, MCP Registry, MCPB, and Smithery delivery pipeline.

API Reference