Features
Everything below ships in Annotide. The free Community edition has every annotation feature for up to three people; Team adds people, and Business adds single sign-on, SCIM, folder permissions, full audit history, quality control, SharePoint and the ML-platform integrations.
Annotation tools & media types
Images, video, audio, text, PDF, time series and LLM evaluation, each with the tools that fit the data.
- Images: box, rotated box, polygon, polyline, point, keypoints, brush/eraser masks, superpixels, measure — plus a smart polygon that draws the outline from a click or a box using a registered model, or the bundled SAM 2.1 image on CPU.
- Video: boxes as keyframes on a track, interpolated in between.
- Text: span selection per class, with relations linking two spans.
- PDF: boxes and text spans per page, with OCR for scanned pages.
- Audio: waveform segments with speaker and transcript.
- Time series: one lane per channel, intervals across all channels or just the ones shown.
- LLM evaluation: rank or pick the better of candidate responses, rate per response, per turn, or for the whole conversation.
- Companion context views (a caption, a second image, the source text) sit alongside the item being annotated.
Quality control
A review workflow with consensus and gold items, and a dashboard that shows where quality stands.
- Review queue: approve, reject with a comment, or correct the result yourself before approving — every save is a new version, nothing is overwritten.
- Consensus: route an item to several annotators and compare their results.
- Gold items: known-answer items that measure annotator accuracy.
- Bulk actions on the item grid — approve, return to queue, assign, tag — with a reason listed for anything they don't apply to.
- Quality panel: agreement between annotators and accuracy against gold items, alongside progress and throughput.
Models & active learning
Pre-label with your own models, prioritize the queue by uncertainty, and track model lineage.
- Pre-labelling: run a registered model over new items; its suggestions arrive as drafts people correct — never a replacement for human review.
- Active learning queue order: prioritize annotation by prediction uncertainty, so the least-confident predictions reach a person first.
- Model family view: distillation and quantization lineage between model versions, in one readable view.
- Managed identity and service principal support for model endpoints, so keys don't have to sit in configuration.
- Bundled SAM 2.1: an optional model image for the smart polygon tool, CPU only, weights baked in — no PyTorch, no GPU, nothing to download at runtime.
- External producers: register a model with no endpoint and post its pre-labels yourself — how the MCP server lets an AI agent pre-label a project.
Data & exports
Snapshot a dataset, split it, and export or import in the format your training pipeline expects.
- Snapshots freeze a filtered, versioned dataset, optionally split into train / val / test by group.
- Exports: COCO, YOLO, YOLO-pose, spaCy, CoNLL, LLM preference pairs, segments (RTTM), or Annotide's native JSON.
- Imports: COCO, YOLO, VOC, CVAT, Label Studio, with a dry run that reports what would happen first.
- Dataset filters by item/annotation status, path, class, annotator or date, so an export is exactly the slice you mean.
- Results are published to your result storage as they happen, through an outbox — not just at export time.
Integrations, SDK & MCP
A typed Python SDK and CLI, an MCP server for AI agents, and webhooks into the systems you already run.
- Python SDK and CLI (pip install annotide): list projects, take snapshots, export, import, wait on jobs — scriptable and CI-friendly.
- MCP server (annotide mcp) lets an MCP client such as Claude Code or Claude Desktop pre-label a project with a scoped service-account key.
- Webhooks deliver signed JSON to your own systems, or post to Slack / Microsoft Teams.
- Retraining events notify your ML pipeline (MLflow, Databricks, Azure ML) when a snapshot is ready, and record whatever model version comes back with its lineage.
Security & SSO
OIDC and SAML single sign-on, SCIM provisioning, an audit log, and telemetry you can see and turn off.
- Single sign-on over OIDC or SAML 2.0 — Microsoft Entra ID, Okta, Google Workspace or any provider — with group-based role mapping.
- SCIM 2.0 user provisioning.
- Per-project roles (annotator, reviewer, owner, viewer), optionally limited to a subset of a project's folders.
- Audit log of sign-ins, project changes, exports, membership changes and review verdicts.
- API keys and service accounts scoped to exactly what they need to do.
- Telemetry is a weekly heartbeat, on by default, can be turned off, carries an install id, counts and salted hashes — never names, e-mail addresses or media.
Deployment
Docker Compose for evaluation, a Helm chart for production on any cloud, and a ready-made Azure path.
- Docker Compose: the whole stack on one machine, for evaluation, demos and development.
- Helm chart: API, worker, frontend and an optional reference model service on any Kubernetes cluster, on premises or in any cloud.
- Azure Terraform module: AKS, PostgreSQL, Redis, storage and Key Vault, using only free networking controls — no private endpoints, private DNS, NAT gateway or VPN to pay for.
- English and Finnish UI.