# Video

> Analyze a video (Loom, YouTube, Vimeo, TikTok, Instagram, X/Twitter, Twitch, Dailymotion, Facebook, direct URL, or local file) — transcript, key frames, OCR text, metadata, annotated timeline — and answer questions about it with timestamps.

- Skill: `guimatheus92/video` (Agent Skill)
- Install (CLI): `npx skillmds@latest add guimatheus92/video`
- Raw SKILL.md: https://api.skillmd.com/api/skills/guimatheus92/video/raw
- Safety review: pending (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- License: MIT
- Author: guimatheus92 (https://skillmd.com/u/guimatheus92)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/guimatheus92/video

---


Analyze the given video and answer the user's question (or summarize it if no question was asked). Always cite timestamps (`M:SS`) in your answer.

## Route A — video-analyzer MCP tools available (preferred)

If the `video-analyzer` MCP server is connected in this session, call its tools directly — do not use the CLI:

- General question or no question → `analyze_video` (detail `"standard"`)
- "What happens at X:XX" / a specific moment → `analyze_moment` (time range) or `get_frame_at`
- Question answerable from speech alone → `get_transcript` (fast, no download)
- Title / duration / views / comments only → `get_metadata` (no download)
- Motion or fast UI changes → `get_frame_burst`

**Dense UI capture** (terminal, dashboard, IDE, spreadsheet — the meaning is in small text): pass `maxWidth` on any of these tools. Emitted frames are capped at 800 px wide by default, which turns a 1920×1080 screencast into 800×450 and drops a 15 px UI font below what a vision model can read. `maxWidth: 0` keeps the source resolution; a value like `1568` is the middle ground. Native frames cost several times more context, so raise it for the close read, not for the overview.

## Route B — no MCP server (any agent with a shell)

Run the one-shot CLI via Bash (first run downloads the npm package — slow is not broken; progress streams on stderr):

```bash
npx -y mcp-video-analyzer@latest analyze "<video-url-or-path>"
```

stdout is a single JSON document: `metadata`, `transcript` (timestamped entries), `ocrResults` (on-screen text), `timeline`, `warnings`, and `frames` — an array of `{ time, filePath, mimeType }` pointing to JPEG key frames on disk. Then:

1. Parse the JSON from stdout.
2. Read the `frames[].filePath` images (in parallel) when the question needs visuals.
3. Answer from transcript + OCR + frames, citing timestamps.

Useful flags: `--detail brief|standard|detailed` (brief = metadata + transcript only, no frame extraction — the fast/cheap path), `--fields metadata,transcript` (filters the emitted JSON only; frames are still computed at standard detail), `--max-frames <1-60>`, `--max-width <px>` (frame width cap, default 800; `0` keeps source resolution — use it for dense UI captures whose payload is small text), `--language <code>` (force transcription language), `--out <dir>` (where frames are copied), `--force-refresh`. Run `npx -y mcp-video-analyzer@latest analyze --help` for the full list.

## Prerequisites & degradation

- Node.js 22.12+ (required). `ffmpeg` is bundled — no install needed.
- Platform URLs (YouTube, Instagram, TikTok, …) require `yt-dlp` on PATH; direct `.mp4/.webm/.mov` URLs and local files work without it. Loom transcript, metadata, and comments need no `yt-dlp` either. Loom **frames** usually do — Loom serves most videos as separate DASH video+audio streams that only `yt-dlp` fetches and merges; a CDN fallback covers some videos without it.
- The tool never fails on partial results: the `warnings` array carries actionable hints (yt-dlp install, `YTDLP_COOKIES_FROM_BROWSER` for Instagram/age-restricted, missing Whisper backend). Relay relevant warnings to the user instead of treating them as errors.
- An empty transcript alongside a "silent audio" warning means the video genuinely has no speech (common for muted Reels/Stories) — that is content, not a failure.
- Only `http(s)` URLs to **public** addresses are accepted. A URL on `localhost`, a private/LAN range, a `.local` name, a UNC share, or a non-HTTP scheme is rejected up front with a message saying so — that is a deliberate refusal, not a transient error, so do not retry it or try to work around it. When the user really does want a video from their own network, tell them to restart the server with `MCP_ALLOW_PRIVATE_URLS=1`. Cloud metadata endpoints stay blocked even then.