Label Studio

Heartex built it, HumanSignal sells the commercial editions, and the core has stayed Apache 2.0 the whole way. Coverage is the distinguishing property: images, video, audio, text, HTML, time series and combinations of them, each described by a labelling config that reads like small XML. Change the config and the interface changes with it. That single mechanism is why one deployment can serve a vision team and an LLM evaluation team without running two products.

The ML backend is the piece worth understanding. Wrap a model behind a small HTTP interface, Label Studio calls it for predictions, and the annotator corrects a prediction instead of starting from an empty canvas. Corrections flow back for retraining, which is the active-learning loop without a bespoke framework. Storage connectors point at S3, GCS or Azure Blob so images sync by reference, which stops mattering as a convenience and starts mattering as a requirement once a dataset passes a terabyte. Output is one JSON shape whatever went in.

The free-versus-paid line is drawn at governance, not at labelling. Community gets the full annotation experience and none of the team machinery: no role-based access control, no consensus or agreement metrics, no reviewer stage, no dashboards, no SSO, no audit logs. Starter Cloud is $99 a month with $49 per additional user up to twelve, Enterprise is quoted. A solo engineer or a pair loses effectively nothing by staying on Community. A twenty-annotator operation hits the ceiling within a week.