Frequently Asked Questions

An open source media asset platform. You point it at your files and it transcribes, detects, describes and indexes what is inside them, then gives you that back as an archive you can search, browse and drive from an API. It is two cooperating pieces: Loom, the asset management server and its web UI, and Cortex, the processing engine that runs the pipelines. The documentation starts with both.

No. The version in the tree is 1.0.0-SNAPSHOT, there are no published artifacts, and interfaces can still move. What does exist is a demo container that runs today — see Getting Started — and a written scope for 1.0.0. Treat everything you read here as a description of a project under active development rather than of a shipped product.

One developer: Johannes Schüth, in Vienna. The team page has the longer version. Contributions and questions are welcome in the Discord or on GitHub.

MetaLoom itself is Apache License 2.0, which permits commercial use. The models you run it with are a separate question: MetaLoom ships no model weights at all — every model is a configuration value you supply — and two of the defaults are non-commercial, namely the face model packs and the default image-generation model. Model Licenses maps every node to the model it loads by default, its licence and whether that licence allows commercial use, and gives a permissive-only configuration.

Two things. It runs on your own hardware, so the media never has to leave it. And the processing is a pipeline you assemble yourself out of nodes — transcription, face detection, captioning, deduplication, translation and so on — rather than a fixed feature set. You can try that model in the browser without installing anything, in the pipeline editor.

For a first look: Docker, and one command — the demo image bundles the server, its database and sample content (Getting Started). For a real deployment: PostgreSQL, the Loom server and at least one Cortex worker. The Docker playbook builds that on a single host and the Kubernetes playbook does it in a cluster.

Not to run MetaLoom, and not for the whole node catalogue — hashing, metadata extraction, thumbnails, format inspection and deduplication are ordinary CPU work. The AI nodes are where a GPU matters: speech-to-text, vision-language captioning, object and face detection, depth maps, image and video generation. Several of those talk to a model server or a sidecar you host, so which hardware they need is decided by the model you point them at, not by MetaLoom.

Only if you configure a node to send it somewhere. MetaLoom has no hosted service, no telemetry and no phone-home. Nodes that call a model do so against the endpoint you configure, which is normally one you run yourself; the speech-to-text node runs the model locally in-process. The cloud source nodes are the deliberate exception — they exist to pull files in from Google Drive or OneDrive, so they talk to those services by definition.

Point a filesystem source at a folder, upload through the UI or the API, or pull from S3, Google Drive or OneDrive with the matching source node. Files can be referenced where they already are rather than copied — see Binary Storage.

Yes — everything the UI does goes through the same public API. There is a REST API with a generated OpenAPI document, a GraphQL API, clients for Java and Python, a command line tool and an MCP server for driving MetaLoom from an AI coding agent.

Depends on which nodes you put in the pipeline. Speech becomes a transcript; faces become detections that can be clustered into people; objects, scenes, dominant colours, depth and image quality all become stored properties; documents and images give up their embedded metadata and text; and a vision-language model can describe or caption a frame in prose. The node catalogue is the full list, one page per node, each with a picture of the node running.

Video, images, audio and documents. Format handling is delegated to the libraries that do it best — FFmpeg and OpenCV for media, Apache Tika for the long tail of document and container formats — so the answer is close to “what those read”. Individual nodes state what they accept on their own page.

Yes. Loom indexes text and embeddings and can answer similarity queries — the near-duplicate video review workflow is built on the same machinery. See Search Indices. The built-in chat agent puts a natural-language front end on it.

Yes, and it is a supported path rather than a fork. A node can be written in Java (examples) or in Python (playbook), and the script node runs a small snippet inline without building anything at all.

Not unless you ask it to. Analysis nodes only read. The nodes that write — the ones that move, watermark or transform a file — are separate, named for what they do, and deduplication in particular never moves anything until a person has confirmed the match.

A proposed commercial edition, built on top of the open source platform for teams who would rather not run it themselves. It is a plan, not a product: see MetaLoom Studio.

No. Nothing that is open source today moves behind the commercial edition, and Studio does not meter processing. Those two promises are written into the plan on purpose, and this page is one of the places they are kept.

Nothing has been priced yet, because nothing has been released yet. Pricing will be announced with 1.0.0.

No. There is no analytics of any kind, no cookies set by this site and no tracking scripts. It is static files on GitHub Pages. What the hosting itself processes — and the one third-party request the site makes, for its webfonts — is written down in the Impressum.

It runs in your browser. Every documentation page is embedded at build time and the search box downloads that index plus a 7 MB model the first time you focus it, then matches locally — both literally and by meaning. Nothing about a search is sent anywhere, because there is nowhere to send it to.

Partly, and the project says which parts and since when rather than leaving you to guess: see AI Disclosure.