io.github.KitDevUA/video-vision-mcp

video-vision-mcp

An MCP server that gives Claude Code the ability to analyze any video — a local file or a URL — through one set of tools.

Claude can't watch video natively (only text + the first frame of an image). This server converts a video into sampled frame images + an audio transcript, or — when a Gemini key is present — a native Gemini analysis of the whole video.

It is standalone: give it a ready video (a local path or a direct URL) and it does the rest. It does not connect to Jira/Slack/etc. If a video lives behind an integration, fetch it with that integration first (download to a file or get a direct URL), then hand the file_path or url to this server.

Scenario: a Jira bug ticket has only a screen-recording, no text. Your Jira MCP downloads the attachment to a temp file → analyze_video file_path=/tmp/bug.mp4 → you see the frames + transcript (or Gemini's analysis) and can reason about the bug.

Three backend tiers (auto-selected)

Precedence: Gemini > OpenAI > Groq > local. Set VIDEO_MCP_DISABLE_GEMINI=true to force tiers 1/2 even with a Gemini key. The backend used is named in every result.

Privacy: tier 1 never uploads anything. Tiers 2/3 print a one-time notice in the session the first time video content is sent to a third party.

Tools

• analyze_video — frames + transcript + metadata (the main tool). frame_interval sets seconds between frames (default 1.0; e.g. 0.5/0.25/0.1 denser, 2/5 sparser).
• get_video_transcript_only — transcript text only.
• extract_frames_at — frames at specific timestamps ("00:42", "1:05", 12.5).
• list_recent_analyses — cached analyses + backend used.

Install

Requires Python ≥ 3.10. A single install pulls everything — backends, plus the ffmpeg and whisper.cpp dependencies. Nothing is ever installed globally on your machine (no brew/apt/winget, no sudo).

Use it (recommended)

With uv you don't install it explicitly — uvx runs the published package on demand (see Register in Claude Code). To install into an environment instead:

uv pip install video-vision-mcp     # or: pip install video-vision-mcp

From source (development)

git clone https://github.com/KitDevUA/video-vision-mcp.git
cd video-vision-mcp
uv venv && source .venv/bin/activate
uv pip install -e ".[dev]"          # all backends bundled

Dependencies — fully self-contained

• ffmpeg / ffprobe: if they are already on your PATH, those system binaries are used. Otherwise the bundled static-ffmpeg package supplies them (fetched once into its own local cache — never a system-wide install).
• whisper.cpp (tier 1 transcription): shipped as the bundled pywhispercpp binding (prebuilt wheels; builds from source only if no wheel exists for your platform/Python). A whisper-cli already on PATH is used if present.
• whisper model: the ggml model (base by default) downloads from Hugging Face into the cache on first transcription. Override with VIDEO_MCP_WHISPER_MODEL (tiny/base/small/medium/large-v3) or VIDEO_MCP_WHISPER_MODEL_PATH.
• cloud-only: set OPENAI_API_KEY / GROQ_API_KEY (tier 2) or GEMINI_API_KEY (tier 3); whisper.cpp is then never invoked.

Configure

cp env.example .env
# edit .env — nothing is required for tier 1

See env.example for every variable — all optional (API keys and tuning). Tier 1 needs none.

Register in Claude Code

Add to your project .mcp.json (or global config) — see .mcp.json.example:

{
  "mcpServers": {
    "video-vision": {
      "command": "uvx",
      "args": ["video-vision-mcp"],
      "env": { "VIDEO_MCP_ENV": "/abs/path/to/.env" }
    }
  }
}

uvx downloads and runs the published package automatically — no manual install step. VIDEO_MCP_ENV is optional (tier 1 needs no keys); point it at your .env if you use the cloud backends. For local development against a checkout, use "args": ["--from", "/abs/path/to/video-vision-mcp", "video-vision-mcp"] instead. Restart Claude Code; the video-vision tools then appear.

Cache

Results are cached at ~/.cache/video-vision-mcp/ keyed by (file hash, backend, frame interval) — re-analyzing the same video is instant, and switching backends or intervals keeps each result separately. Downloaded URLs and whisper models live under the same dir. Override with VIDEO_MCP_CACHE_DIR.

Cached analyses and downloaded videos older than VIDEO_MCP_CACHE_TTL_HOURS (default 24) are pruned on startup and skipped on read; set 0 to keep them forever. Whisper models are never pruned (expensive to re-download).

Using it with an integration (e.g. Jira, Slack)

This server is deliberately standalone — it never talks to Jira, Slack, or any other service. When a video lives behind an integration, let that integration's MCP fetch it, then pass the result here:

1. The integration MCP downloads the attachment to a local file (or gives a direct, publicly reachable URL — an authenticated API URL won't work with url).
2. Call analyze_video file_path=<downloaded file> (or url=<direct link>).

This keeps auth and service-specific logic where it belongs, and lets one video tool serve every source.