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

For subsecond motion, use fractional bounds (for example, 0:01.1 to 0:01.3). FFmpeg burst extraction returns up to the requested count of real source frames; it does not duplicate frames when the requested rate exceeds the source rate. Preserve fractional timestamps when supplied. Browser fallback uses fractional seek targets, but those are not measured source-frame presentation timestamps.

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.

get_frame_burst returns a JSON text block containing { frameCount, from, to, warnings, frames: [{ time, timingOrigin }] }, followed by images in exactly the same order as frames. timingOrigin: "source-pts" means measured video-relative FFmpeg presentation time; "seek-target" means requested browser seek time. Empty/degraded results contain frames: [].

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):

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.