LabelStudio

Label Studio's core bet is that annotation interfaces are a programming problem, not a template-picking problem. The XML tag system has been there for years, but the Vibe Code Interfaces Agent is what makes that bet accessible: you describe the interface in natural language, and Select Element mode lets you click the precise element in the live preview so an edit targets that element rather than a described one. Voice input with streamed reasoning and tool activity rounds it out for people who'd rather talk through a change. Where it earns its keep is breadth under one roof. Image classification, object detection with boxes, polygons, circular and keypoint shapes, frame-by-frame tracking, and semantic segmentation cover vision. On the document side, DocLang gives you a layout template with reading order, in-browser OCR and XML preview, and the Docling integration ships a purpose-built interface with live IBM watsonx predictions. Audio handles transcription over waveform or spectrogram, speaker diarization and emotion recognition. Time series covers classification, segmentation and event recognition, and multi-modal projects pair a call recording's transcript against its audio or an image against its OCR text. VideoVector adds SAM 2 segmentation and tracking across frames for video work. The evaluation surface is the part competitors undersell. LLM evaluation with custom benchmarks, rubric scoring and side-by-side comparison; RAG and retrieval QA that grades answers against source documents and scores retrieval relevance; RLHF preference and ranking collection; and agentic trace review that connects observability tools so a human can sit in the loop on agent runs. The July 2026 rationale post on rubric design — that clear guidelines still produce disagreeing annotators unless rubrics are explicit, inspectable and testable before labeling starts — is a fair summary of what the product is trying to solve at that layer. The honest constraints. ML backends and custom integrations need Python, Docker and often Kubernetes; that's real engineering time, not a checkbox. Real-time collaboration at scale generally means a Kubernetes deployment, not a single container. Service accounts are limited to one per organization by default. DICOM (September 2026) and LiDAR (September 2026) interfaces, the Enterprise Vibe Code backend, SSO/SAML, SCIM seat provisioning including View-Only seats, RBAC, SLAs, configurable reject actions and annotator pause modes are Enterprise-scoped. The community edition gives you the platform and leaves deployment, upgrades and annotator QA tooling to you. Against alternatives: Supervisely and Dataloop give you a managed environment and fixed interface surfaces — easier to start, less shaped to your data. Label Studio asks for an engineering owner and returns control over storage, model wiring and evaluation criteria. If you have that owner, it's the better long-run fit for teams that annotate more than one modality and also