MetaLoom 1.0.0 | MetaLoom

MetaLoom 1.0.0 has not been released yet. This page describes the scope of the upcoming release so you can judge whether it fits your setup. No 1.0.0 container images or Maven artifacts are published — everything currently builds from source at 1.0.0-SNAPSHOT. The release date is not fixed; follow the Discord or the blog to hear about it first.

MetaLoom 1.0.0 is the first release intended for production use. It marks the point where the building blocks — the server, the processing fleet, the APIs and the front ends — line up into one platform: point it at media, and it works out what is in it, then stores, indexes and serves that back to you.

What Is In the Release

Pipeline engine

Loom owns the pipeline graph. It persists run state, resolves the DAG and dispatches individual source, node and segment tasks to registered Cortex workers over a WebSocket. Workers hold no database and run one task per slot, so scaling out is a matter of starting more of them — and heavy kinds (speech-to-text, face detection, vision models) can be pinned to GPU hardware. Pipelines are versioned, can be re-run, and every node result is written back with the producer version that created it.

See Pipeline and Operation.

Processing nodes

A catalogue of built-in node kinds covering hashing and perceptual fingerprinting, deduplication and consistency checks, thumbnails and contact sheets, face detection and face descriptions, OCR, document extraction (Tika), speech-to-text (whisper.cpp), scene detection, captioning, vision-language models, LLM prompting, text-to-speech, quality metrics, filesystem and asset sources, and routing filters. Custom kinds can be written in Java or as a Python worker speaking the same wire protocol.

See Nodes and the custom node playbook.

Chat and AI agent

A server-side agentic loop that answers questions about your assets in natural language — with per-chat Skills, a persistent Memory bank, publishable Sessions and an optional isolated coding sandbox for agent-authored scripts.

See Chat & AI Agent.

APIs

REST — roughly 130 endpoints across 35 resource groups, with a generated OpenAPI document, an in-browser API explorer and a typed Java client.
GraphQL — nested reads in a single round trip, with an embedded GraphiQL explorer.
WebSocket — live task and pipeline events, plus the processor protocol Cortex workers speak.

See REST API, GraphQL API and Java Client.

Front ends

The Loom UI (React) ships inside the server image and covers assets, libraries, tags, tasks, detections, pipelines, monitoring, ACL and the chat agent. Loom App wraps the same UI in an Electron desktop window. The metaloom CLI talks to the same REST API and ships as a GraalVM native binary as well as a runnable JAR.

See Loom UI and Command Line.

Storage, search and access control

Assets can be referenced in place on disk or uploaded and stored content-addressed. Metadata is indexed for full-text and vector search (Elasticsearch, Lucene, Qdrant). Permissions are granted on object level through users, groups and roles, and external identity providers (Keycloak, Auth0, Okta) are supported over OAuth2 alongside built-in JWT auth.

See Loom Features and Authentication.

Deployment

Container images for the server, the demo, the Cortex worker and the agent session runner — each of the server images in a JVM and a GraalVM native variant. Two Helm charts (helm/loom, helm/cortex) deploy the stack on Kubernetes. Prometheus metrics are exposed by both Loom and Cortex.

See Container Images, Helm Charts and the Kubernetes playbook.

Important Points

Topic What you should know

Topic	What you should know
Release status	Not released. The version in the source tree is `1.0.0-SNAPSHOT`; published artifacts will follow when the release is cut.
License	The platform is Apache 2.0. That covers the code — not the models you point it at.
Models are not shipped	MetaLoom ships no model weights. Every model is a configuration value (a path, a repo id or an endpoint), so the license that applies is the one of the model you supply.
Two components are non-commercial	The InspireFace model packs used by face detection are academic-use only (which also affects face descriptions downstream), and Ideogram 4.0, the model behind the planned image generation node, is non-commercial. See Model Licenses before deploying commercially.
Database	PostgreSQL is the production backend (jOOQ + Flyway). In-memory and filesystem backends exist for development and tests.
Hardware	CPU-only works. Speech-to-text, vision-language and LLM nodes are much happier with a GPU — pin those kinds to dedicated workers.
Node availability per worker	A worker executes the node kinds its build registers. Check the node catalogue and the worker’s `/api/v1/processors` entry rather than assuming every kind is present everywhere.
API stability	1.0.0 is where the REST and GraphQL surfaces become something we keep stable. Until it is cut, endpoints may still move.

Release status

Not released. The version in the source tree is 1.0.0-SNAPSHOT; published artifacts will follow when the release is cut.

License

The platform is Apache 2.0. That covers the code — not the models you point it at.

Models are not shipped

MetaLoom ships no model weights. Every model is a configuration value (a path, a repo id or an endpoint), so the license that applies is the one of the model you supply.

Two components are non-commercial

The InspireFace model packs used by face detection are academic-use only (which also affects face descriptions downstream), and Ideogram 4.0, the model behind the planned image generation node, is non-commercial. See Model Licenses before deploying commercially.

Database

PostgreSQL is the production backend (jOOQ + Flyway). In-memory and filesystem backends exist for development and tests.

Hardware

CPU-only works. Speech-to-text, vision-language and LLM nodes are much happier with a GPU — pin those kinds to dedicated workers.

Node availability per worker

A worker executes the node kinds its build registers. Check the node catalogue and the worker’s /api/v1/processors entry rather than assuming every kind is present everywhere.

API stability

1.0.0 is where the REST and GraphQL surfaces become something we keep stable. Until it is cut, endpoints may still move.

After 1.0.0

On the roadmap, not in this release: an S3 storage backend, selective import/export between instances, server-side image manipulation (resize, crop, focal point), a gRPC surface and the imagegen node.

Try It Before the Release

The demo container runs the whole thing on a laptop with seeded assets, pipelines and users:

docker run -p 8092:8092 metaloom/loom-demo:latest

Then follow Getting Started. If something does not behave the way this page describes, that is worth telling us about on Discord — feedback before a 1.0 is worth more than after it.

All announcements