Sometimes the easiest way to describe a file is what is in it, not what it is called. A folder might contain the image you remember, but its filename is not necessarily going to help you find it.
Media Indexer is a local-first tool for searching image and video folders in plain English. It combines a command-line workflow for indexing with a browser interface for exploring the results. The media and its search index stay on your machine.
The workflow
The guided command lets you choose folders, build an index, and start the local web interface. Once the index is ready, a query can describe a scene, a subject, or a visual style. Examples include “sunset over water” or “warm editorial portrait.”
The browser also accepts an image for visual similarity search. This gives the archive two entry points: describe what you want, or show it something similar.
The tool handles common image formats and video formats including MP4, MOV, MKV, and WebM. It scans folders recursively and creates thumbnails so results can be browsed without repeatedly opening the original files.
How search works
The indexer generates CLIP embeddings locally. These give text and images representations that can be compared for similarity. A text query can therefore retrieve an image without requiring its filename to contain the same words.
Each indexed asset also has local metadata: its path, dimensions, timestamps when available, a thumbnail, labels, and its search embedding. File hashes or stable video-frame signatures help identify content that has already been processed.
This matters when an archive changes gradually. Reindexing skips unchanged files instead of repeating all the work every time a folder is scanned.
Making video searchable
Indexing every frame of every video would create a lot of redundant work. The project uses sparse representative frames instead.
A match from a video retains the source path and the timestamp of the matched frame. That makes the result more useful than a thumbnail detached from the clip it came from.
There is an important limit to this approach. Matching a frame does not mean understanding the entire clip. A short action between sampled frames can be missed, and a still image does not capture all the temporal information in a video. Sparse indexing is a practical compromise for this tool, not full video understanding.
Keeping the archive local
The application stores its index, thumbnails, vectors, and cached frames in the user’s app-data directory. It does not modify the original media files.
No account, cloud media upload, or external inference API is required for the search workflow. The first run may download model weights, which are then reused locally. That distinction is useful: local inference can still need an initial model download.
The local web interface is served at port 8000. The CLI also provides separate commands for indexing a folder, serving an existing index, checking status, and reindexing configured folders.
Current limitations
There is still work to do for very large archives. The project also notes that deleted files are not fully cleaned out of the index automatically yet.
The automatically generated labels are another convenience rather than a complete description of an asset. CLIP-based labels are less detailed than descriptions from a larger vision-language model. They help with browsing, but they should not be mistaken for an exhaustive account of what a file contains.
Try it
After installing the project and FFmpeg for video support, the guided workflow starts with:
media-indexer run
For an archive that is already configured:
media-indexer reindex
media-indexer serve
The repository documents installation, supported commands, storage locations, and development tests.