Hosting

Local Python MCP server

Transport

stdio

Endpoint

Local process; no hosted endpoint documented

Authentication

No API key for local use or public YouTube videos; local-file access is limited to configured directories.

Requirements

• Python 3.11 through 3.13 and uv for the documented locked install.
• FFmpeg and FFprobe 6 or newer.
• Local storage for video, frame cache, and an explicitly installed faster-whisper model when transcription is needed.

Setup steps

1. Clone the maintainer repository and run `uv sync --frozen`.
2. For speech recognition without suitable captions, run the virtual-environment `video-mcp install-model base` command, then `video-mcp doctor`.
3. Point Codex's local stdio MCP configuration at the absolute virtual-environment `video-mcp` executable with `serve`; add `--allow-dir` roots if inspecting local files.
4. Inspect a public YouTube URL, prepare the selected range, read timeline pages, preview frames, then retrieve the frames needed for close inspection.

Nine stored tools are documented against the matching maintainer repository.

• inspect_youtube_video

  Check one public YouTube video's metadata, range, and preparation estimates.

  Parameters: url, range (optional), profiles (optional)
  01
• inspect_local_video

  Check an allowed local video's metadata and preparation options.

  Parameters: path, inspection_mode, range (optional)
  02
• prepare_youtube_video

  Acquire a bounded YouTube video and prepare transcript, frames, and timeline.

  Parameters: url, range (optional), sampling (optional), transcript (optional), images (optional)
  03
• prepare_local_video

  Prepare an allowed local video with the same processing controls.

  Parameters: path, range (optional), sampling (optional), transcript (optional)
  04
• read_video_timeline

  Read paged transcript and frame events from a prepared job.

  Parameters: job_id, time range (optional), continuation token (optional)
  05
• get_video_frame_previews

  Return small image previews for selected frames or time ranges.

  Parameters: job_id, frame IDs or time range, count (optional)
  06
• get_video_frames

  Return selected stored full frames as MCP images.

  Parameters: job_id, frame IDs, maximum count (optional)
  07
• delete_video_job

  Remove one prepared job and its previews.

  Parameters: job_id
  08
• video_mcp_doctor

  Report dependency, model, cache, and processing-limit health.

  09

• Version 0.1.0 is an initial release; the maintainer says non-Administrator Windows production stdio startup has not had a live acceptance test.
• A sampled frame timeline cannot prove every brief visual action or describe music and sound design.
• Public YouTube acquisition accepts one canonical video URL and deliberately rejects playlists; jobs and output size are bounded.

• maintainer repository readme

  Documents nine public tools, setup, transcript policy, frame sampling, Windows caveat, and job limits.

• creator github avatar

  Local copy decoded as a 420 by 420 PNG and visually reviewed.

• maintainer repository documentation

  Maintainer repository describes local-first transcript and frame-based video understanding; existing nine-tool inventory was already sourced from this repo.

Documentation checked 2026-10-28T00:00:00Z. Connection has not been runtime tested.