transcriptor-mcp

Each tool that takes a video accepts url. This is a link from a supported platform or a plain YouTube ID. Each tool returns content (text for the chat) and structuredContent (typed JSON for your code).

get_transcript

Clean plain text, without timestamps, HTML, or speaker names. Without lang, the tool returns the track in the video's original language. Most platforms other than YouTube do not say which language a video is in; when the tool cannot tell which track that is, it answers with the list of tracks, and you call it again with type and lang. The inputs are the same as for get_raw_subtitles.

Response: videoId, url (the video page, as the server resolved it), type, lang, text, is_truncated, total_length, start_offset, end_offset. When more text is available, the response also has next_cursor.

get_raw_subtitles

Raw SRT or VTT content, in parts.

Input:

• type — official or auto. Without lang, the tool picks a track of this type
• lang — a language code or track name, as get_available_subtitles lists it. Without it, the video's original language, as for get_transcript
• response_limit — default 50000, minimum 1000, maximum 200000
• next_cursor — the cursor of the previous response

Response: the fields of get_transcript, plus format (srt or vtt) and content.

get_available_subtitles

Response: official and auto. Each field is a sorted list of language codes. Use this tool first, then give type and lang to the tools above.

get_video_info

Extended metadata from yt-dlp:

• identity — videoId, title, description, webpageUrl
• author — uploader, uploaderId, channel, channelId, channelUrl
• numbers — duration, uploadDate, viewCount, likeCount, commentCount
• classification — tags, categories, liveStatus, isLive, wasLive, availability
• images — thumbnail and thumbnails

get_video_chapters

Response: chapters. Each item has startTime, endTime, and title. When the video has no chapters, the list is empty.

get_video_frame

Input:

• timecode — "MM:SS" or "HH:MM:SS.mmm"
• seconds — an alternative to timecode. Give one of the two, not both
• format — jpeg (default) or png
• width — default 1280, maximum 1920, never larger than the source
• quality — 2 to 31, for jpeg only

Response: an image block, plus url, timestampSeconds, timestamp, mimeType, sizeBytes, and width. This tool needs ffmpeg. The Docker image includes it.

get_playlist_transcripts

Input:

• url — a playlist URL, or a watch URL with list=
• type, lang, format — the same as get_raw_subtitles, except that lang is required: the original language is picked only for one video at a time
• playlistItems — a yt-dlp -I value such as 1:5, 1,3,7, or -1
• maxItems — the maximum number of videos

Response: results. Each item has videoId and text.

search_videos

Input:

• query — the search text
• limit — default 10, maximum 50
• offset — the number of results to skip
• uploadDateFilter — hour, today, week, month, or year
• response_format — json (default) or markdown

Response: results. Each item has videoId, title, url, duration, uploader, viewCount, and thumbnail.

Transport. The server accepts POST /mcp only. GET and DELETE return 405. The server is stateless and sends no Mcp-Session-Id.

The Node process does not check bearer tokens. Put a reverse proxy or a gateway in front of it for authentication and TLS. The hosted endpoint works this way.

REST API. A second image gives the same extraction over plain HTTP:

docker run --rm -p 3000:3000 artsamsonov/transcriptor-mcp-api:latest

The Swagger interface is at http://localhost:3000/docs. For a full stack with the API and the MCP server, read docker-compose.example.yml.

Development.

npm ci
npm run build
npm run start:mcp        # stdio
npm run start:mcp:http   # Streamable HTTP on port 4200
npm test

You need Node.js 22 or later (20 still works, but it reached end of life in April 2026), and yt-dlp in your PATH. Frame capture needs ffmpeg, and WHISPER_MAX_DURATION_SECONDS needs ffprobe (both ship in the same package, and in the Docker image). Other scripts: lint, type-check, format, test:coverage, test:e2e:api, and test:e2e:mcp.

Releases. The maintainer cuts them. The steps are in .claude/skills/release/SKILL.md. The version comes from package.json at runtime, through src/version.ts. Pushing a v* tag makes CI build both images and publish the MCP Registry entry from server.json.

Layout. src/mcp.ts (stdio entry), src/mcp-http.ts (Streamable HTTP), src/mcp-core.ts (tools, prompts, widgets), src/youtube.ts (yt-dlp), src/whisper.ts, src/cache.ts, src/index.ts (REST API), load/ (k6), and src/e2e/ (Docker smoke tests).