Clipswarm

What clipswarm MCP server does

clipswarm MCP server connects AI agents to YouTube videos and livestreams for local video clipping. It exposes metadata, transcript, transcript-search, rendering, and result-collection operations over MCP. The rendered output is intended for short-form publishing: the viral style uses a vertical 9:16 composition, a title header, the complete source frame, and animated word-by-word captions.

The server can work with ordinary videos and live broadcasts. Live requests can target the recent live edge or a time that has already aired. For regular videos, captions come from YouTube when available. Live streams and videos without usable captions can use local whisper.cpp transcription, although clips may render without captions when no transcription source is available.

How it works

An agent can first call get_video_info to inspect a video's title, channel, duration, chapters, and live status. get_transcript returns compact timestamped lines and can limit the response to a time range. search_transcript finds phrase matches and returns padded windows suitable for clipping.

The agent then sends one or more jobs to create_clips. Jobs may cover several videos and can be rendered concurrently. The call returns a job ID after roughly 40 seconds rather than waiting for an entire large batch. get_clips retrieves the completed or still-processing results. This job model is useful when an MCP client has a short request timeout.

For livestreams, relative positions such as the last 60 seconds can be used, subject to YouTube's available DVR window. Plain live clips may begin on a prior keyframe unless precise cutting is requested; viral-format clips are frame-accurate.

Setup and configuration

Install Node.js 20 or newer, yt-dlp, and ffmpeg. whisper.cpp is optional for live-stream captions and fallback transcription. The package provides a setup command that checks dependencies and downloads its caption model. Keep yt-dlp current because outdated versions can prevent YouTube downloads.

Run the MCP server with npx -y clipswarm mcp, optionally passing --out to choose the directory for generated clips. The README documents setup for Claude Code, Claude Desktop, Codex, Cursor, VS Code, and Gemini CLI. The server runs locally and does not require API keys or a subscription.

Tools and capabilities

The clipswarm MCP server provides:

• get_video_info for source metadata and live status.
• get_transcript for timestamped, optionally windowed transcript text.
• search_transcript for locating phrases and producing padded clip windows.
• create_clips for parallel rendering across multiple videos.
• get_clips for collecting results from an asynchronous batch.
• Watch-management tools for scheduled discovery, including creating watches, running them, listing watches, and reviewing inbox items.

Rendering options include style, title, captions, accent color, background, label, vertical layout, maximum height, and precise cutting. Watch workflows can search for new videos on a schedule, select moments with AI, replay data, or transcript heuristics, and place items needing agent review in an inbox.

Limitations and notes

The tool depends on yt-dlp and YouTube availability. Live clipping is limited by the stream's DVR window, and a requested end time cannot be in the future. Caption availability varies by source; without YouTube captions or whisper.cpp, output includes a warning and may not contain captions. The repository describes local execution, but it does not provide a hosted-server deployment path for web-only clients.