Build a self-hosted AI video library

Why index a library at shot level?

A folder tells you where a file lives. It rarely tells you where a specific moment is hiding. Metachlorian analyses each video as a sequence of shots and records visual meaning alongside motion, duration, speech, on-screen text and usage rights.

That lets you search for a description such as “slow coastal drone shots at golden hour” and narrow the results by the practical requirements of an edit. You can inspect why each result matched before adding it to a collection.

What you need

For the Docker route, use a machine with Docker and the Docker Compose plugin, Git, and access to your footage. This can be a server, a compatible NAS or your development machine. Keep space available for the library database, downloaded models and generated media.

The default analysis models run on a CPU. A GPU is optional; it can accelerate supported models and enable richer local vision-language captions. The repository documents its CPU measurements and the limits of its evaluations.

Start with Docker

Clone the repository, choose a long administrator password and point FOOTAGE_DIR at your real footage folder. Replace both example values before running:

git clone https://github.com/JeremySNR/Metachlorian
cd Metachlorian

METACHLORIAN_ADMIN_PASSWORD='choose-a-long-password' \
FOOTAGE_DIR=/path/to/footage \
  docker compose -f deploy/docker-compose.yml up -d

Open http://<host>:8765, replacing <host> with the server address, and sign in as admin. On the same machine, the host can be localhost.

The supplied Compose configuration mounts footage read-only and keeps application data and model files in named volumes. Review that configuration when adapting it to your storage environment.

If you prefer a native installation, the repository also documents Python 3.11, uv, model downloads, the frontend build and the local server command.

Make your first search useful

1. Index a representative folder. Start with footage you know so you can judge the results.
2. Describe a moment. Include a subject, setting or remembered line rather than only a filename.
3. Refine your requirements. Duration, shot size and intended use help narrow the selection.
4. Inspect the evidence. Signals carry a source, confidence and model version. Correct mistakes; human corrections take priority after reprocessing.

Add rights before handing footage off

Record the licence, permitted uses, channels, territories, expiry and releases for each file. Individual shots can override the file-level record. When you give search an intended use, Metachlorian filters against those recorded permissions.

The software helps you apply your rights records consistently. Keep those records accurate and complete: it does not independently establish ownership or obtain missing permissions.

Take the selects into your workflow

Build collections and export through OpenTimelineIO, FCPXML or CMX 3600 EDL, or send a package to Cutawan. If you work with AI agents, connect the same library through Metachlorian’s MCP server.

See the repository quick start, Compose configuration and published evaluation for the implementation details.