YouTube 字幕检索与问答MCP服务

yttranscript-mcp

Fetch, clean, semantically search, and ask questions about YouTube transcripts — captions-first (fast, light, rate‑limit resistant) with an optional local Whisper fallback, an MCP server, a CLI, and an async Python library. No YouTube Data API key required. Everything runs locally.

Built for feeding transcripts to LLMs: the clean format strips rolling auto‑caption duplication, HTML entities, markup, and timestamps so you spend the fewest tokens possible.

What makes it special

• Search inside videos. Ask a natural-language question; get the exact timestamped moments with deep-link URLs (https://youtu.be/ID?t=123). Hybrid BM25 + local embeddings, fused with Reciprocal Rank Fusion and diversified with MMR — exa-quality retrieval, fully local, zero new dependencies (it works with no model at all and gets better when you add one).
• Ask questions (local RAG). ytt ask ID "question" retrieves the relevant passages and, if a local LLM is running, writes a grounded answer that cites timestamps. No LLM? You still get the cited passages.
• Cross-video corpus search. Index a library of videos once (ytt index …), then ytt find "query" searches across all of them — exa for your own YouTube collection, in a single SQLite file.
• Beats yt-dlp for transcripts, audio-free. List every caption language (ytt langs), machine-translate captions into any language (--translate), and dump rich metadata + chapters (ytt info) — all from the lightweight captions path, no video download.

vs. the tools you already use

Why it avoids rate limits

• Captions first, audio last. It pulls YouTube's own caption tracks (a few KB) instead of downloading audio. Whisper only runs when no captions exist.
• No watch‑page scrape on the hot path. A hardcoded public Innertube key skips the ~1 MB HTML fetch that triggers bot detection.
• Realistic headers. Every request sends a browser/app User-Agent and language headers (the default python-requests UA is the #1 cause of 429s).
• Multi‑client fallback. Tries ANDROID_VR → WEB → MWEB → watch‑page until one returns captions, sidestepping client‑specific blocks.
• Retry/backoff. Transient 429/5xx are retried with exponential backoff + jitter, honoring Retry-After.
• Escape hatches. YTT_PROXY and YTT_COOKIES_FILE route around IP‑level blocks and age/region gates.
• Caching. SQLite cache avoids re‑fetching.

Install

pip install yttranscript-mcp            # core: captions + CLI + library
pip install "yttranscript-mcp[mcp]"     # + MCP server
pip install "yttranscript-mcp[whisper]" # + local Whisper fallback (needs ffmpeg)
pip install "yttranscript-mcp[whisper,torch]"  # + GPU

The Whisper fallback needs ffmpeg: winget install ffmpeg (Windows) · brew install ffmpeg (macOS) · sudo apt install ffmpeg (Linux).

CLI

ytt transcript VIDEO_ID                     # clean, LLM-ready text (default)
ytt transcript "https://youtu.be/VIDEO_ID"  # URLs work too
ytt transcript VIDEO_ID -f json             # or text | srt | vtt
ytt transcript VIDEO_ID -o out.txt          # save to file
ytt transcript VIDEO_ID --no-whisper        # captions only, never download audio
ytt transcript VIDEO_ID --translate es      # machine-translate captions to Spanish

ytt batch VID1 VID2 VID3                     # many videos, concurrent
ytt search "python tutorial" -n 5            # YouTube keyword search
ytt search "python tutorial" --rank          # NEURAL re-rank by transcript relevance
ytt search "python tutorial" --with-transcripts
ytt cache-stats [--clean]

Semantic search & Q&A (fully local)

# Metadata + chapters, no audio download (beats yt-dlp --dump-json for the essentials)
ytt info VIDEO_ID

# Every available caption language + translation targets (like yt-dlp --list-subs)
ytt langs VIDEO_ID

# Search INSIDE a video — ask a question, get the exact timestamped moments
ytt ask VIDEO_ID "how does the event loop schedule coroutines?"
ytt ask VIDEO_ID "what's the main argument?" --passages-only   # skip the LLM answer

# Build a local corpus index, then search across ALL indexed videos
ytt index VID1 VID2 VID3
ytt find "retrieval augmented generation tradeoffs"
ytt corpus                                   # index stats

ask retrieves the most relevant passages and (if a local LLM is reachable) writes a grounded answer citing timestamps; otherwise it just returns the cited passages. find searches your whole indexed library. Both print deep-link URLs straight to the moment in each video.

Python library

import asyncio
from ytt import get_transcript, search, search_and_get_transcripts

async def main():
    # clean = deduplicated, no timestamps — best for LLMs
    r = await get_transcript("dQw4w9WgXcQ", output_format="clean")
    print(r.source, r.language, r.content[:200])

    for v in await search("python tutorial", max_results=5):
        print(v.video_id, v.title)

    for video, transcript in await search_and_get_transcripts("python", max_results=3):
        if transcript:
            print(video.title, "→", transcript.content[:80])

asyncio.run(main())

Semantic search, RAG, and the cross-video corpus index are all public too:

import asyncio
from ytt import (
    search_in_video, ask_video, index_videos, find_in_corpus, get_video_info, search_ranked,
)

async def main():
    # Search inside one video → timestamped passages with deep-link URLs.
    res = await search_in_video("VIDEO_ID", "the key tradeoffs")
    for p in res.passages:
        print(p.timestamp, p.score, p.url(), p.text[:80])

    # Local RAG: grounded answer + cited passages (falls back to passages if no LLM).
    a = await ask_video("VIDEO_ID", "what is the main conclusion?")
    print(a.answer or a.note)

    # Cross-video semantic search over a local corpus index.
    await index_videos(["VID1", "VID2", "VID3"])
    for p in await find_in_corpus("retrieval augmented generation"):
        print(p.title, p.timestamp, p.url())

    # Neural re-ranking of YouTube results by transcript relevance.
    for video, score in await search_ranked("python asyncio", max_results=5):
        print(f"{score:.3f}", video.title)

    meta = await get_video_info("VIDEO_ID")
    print(meta.title, meta.view_count, "views;", len(meta.chapters), "chapters")

asyncio.run(main())

Embedding backends (how "fully local" stays local)

Semantic search resolves an embedder in this order, and never needs the network by default:

With no embedding model at all, search still works as fast cross-video BM25. Add a real embedder and it automatically upgrades to hybrid lexical + dense ranking. Nothing leaves the machine unless you configure a remote endpoint.

output_format: clean (default for the CLI/MCP), text, json, srt, vtt, summary.

Local summarization (save tokens)

Compress a transcript to a short summary with a local LLM, so the agent that consumes it ingests a summary instead of the whole transcript. Nothing leaves the machine. Opt-in; needs a local model (default Ollama).

ollama serve
ollama pull qwen3.6:27b      # default — smaller: qwen3:8b, qwen3:4b, qwen3.5:2b

ytt transcript VIDEO_ID --summarize
ytt transcript VIDEO_ID --summarize --summary-model qwen3:8b

from ytt import get_transcript
r = await get_transcript("VIDEO_ID", output_format="summary")
print(r.content)   # bullet-point summary

• Long transcripts are summarized map-reduce (chunk → reduce) to fit context.
• YTT_SUMMARY_KEEP_ALIVE keeps the model hot between calls (-1 = forever).
• YTT_SUMMARY_AUTO_PULL=1 pulls a missing model on demand.
• YTT_SUMMARY_PROVIDER=openai targets any OpenAI-compatible server (llama.cpp, LM Studio, vLLM) via YTT_SUMMARY_OPENAI_BASE.

MCP server

yttranscript-mcp          # or: python -m ytt.mcp.server

Claude Desktop / Cursor config:

{
  "mcpServers": {
    "yt-transcript": {
      "command": "yttranscript-mcp"
    }
  }
}

Tools:

• Transcripts: get_transcript, get_transcripts_batch (default to the clean format), summarize_video (local-LLM summary)
• Metadata: get_video_info (title/channel/views/chapters, no download), list_caption_languages (tracks + translation targets)
• Semantic (fully local): search_transcript (timestamped passages inside a video), ask_video (local RAG with cited timestamps), index_videos + find_in_corpus (cross-video search)
• Search: search_videos (add rank=True for neural transcript-relevance re-ranking)
• GPU: setup_gpu, download_cuda

Configuration (environment variables)

Everything is tunable without code via YTT_* env vars:

How it works

video id / url → cache → captions (ANDROID_VR → WEB → MWEB → watch page)
                                  ↓ none?
                          Whisper fallback (download audio, transcribe locally)
                                  ↓
                          cache → format (clean / text / json / srt / vtt)

Semantic search (ask / find / search --rank) runs on top, fully local:

transcript → timestamped chunks (rolling-caption dedup, overlap)
                  ↓
         BM25 (always) + dense embeddings (if a local backend is configured)
                  ↓
         Reciprocal Rank Fusion → MMR diversity → ranked passages w/ deep links
                  ↓ (ask only)
         local LLM writes a grounded answer citing the timestamps

Development

uv sync --extra dev --extra mcp
uv run pytest
uv run ruff check ytt/ && uv run black --check ytt/

License

MIT