Summary

The analysis is stored, and the model reads only what it asks for.

The design decision that makes this practical is that `analyze_video` returns an identifier rather than the content: a long video registers cheaply and the agent then pulls the summary, one phase, or the steps that match a search. Locking provider and model to environment variables is the other good call — an agent cannot quietly switch to an expensive cloud model mid-session.

What it is

A video-understanding server that runs on local models by default. It extracts frames with ffmpeg, sends them to a vision model, and returns a `flow` array with one entry per scene carrying the visible scene, the subjects, the action, on-screen text, persistent context and observations. When the video has audio it transcribes with Apple MLX Whisper and merges the speech into each step. Inputs can be a local file, an image, a PDF, or a URL from YouTube, Loom, Vimeo or Twitter.

What you get

`analyze_video` — run the analysis and get back an `analysis_id` plus metadata, with the full result stored on disk rather than dumped into the conversation
Pull only the part you need: `get_summary` at brief, medium or full depth, `get_step` for one step, `get_steps` for a range, `get_phase` for every step in a named phase, and `get_full_analysis` when you do want everything
`search_analysis` — return only the steps matching a query, so a long video costs you the hits rather than the whole transcript
`get_transcript` — a slice of the speech transcript between two timestamps
Housekeeping: `list_analyses`, `delete_analysis`, `get_usage_stats` for a per-model token and cost breakdown, and `clear_usage_stats`
Analyses also readable as resources at `vidlizer://analyses`, `vidlizer://analyses/{id}` and `vidlizer://analyses/{id}/summary`

Requirements

Nothing when running locally against Ollama — no API key and no data leaving the machine. macOS with Python 3.10 or higher; the package is `vidlizer` (0.2.19 in pyproject) and the MCP entry point is `vidlizer-mcp`, run with `uvx "vidlizer[mcp]" vidlizer-mcp` or installed through pipx with the `[mcp]` extra — a plain pipx install creates the entry point without the `mcp` package, and needs `pipx inject vidlizer mcp`. Provider is chosen with `PROVIDER`: Ollama with a vision model pulled, any OpenAI-compatible server through `OPENAI_BASE_URL` and `OPENAI_MODEL`, or OpenRouter with `OPENROUTER_API_KEY`. ffmpeg is installed via Homebrew on first run if missing. The model and provider are fixed by environment variables and cannot be changed by the calling agent.

Setup effort

One command plus a key — uvx "vidlizer[mcp]" vidlizer-mcp, then supply credentials