Get started
Introduction
What BOAT is, how it is put together, and where to go next.
BOAT, the Best Open Annotation Tool, is a self-hosted image labeling service for individuals and small teams. It runs on one machine you control, stores everything in a local SQLite database and your filesystem, and does not need a database server, container runtime, email service or paid plan.
Vision models are optional add-ons. Install a model such as SAM3 with one command and its tools appear in the annotation editor: type “forklift” to get boxes, click an object to get a mask, or draw one box and let the model find every similar object.
What you can label
BOAT supports seven annotation types, each with its own editor:
| Annotation type | Output | Model tools available |
|---|---|---|
| Object detection | Bounding boxes | Text prompt, box prompt, cascading, auto label |
| Oriented object detection | Rotated boxes | Auto search |
| Instance segmentation | Object masks | Text, box and point prompts, auto search |
| Semantic segmentation | Per-pixel classes | Text, box and point prompts, auto search |
| Crowd counting | Points | Text prompt |
| Keypoint detection | Skeleton keypoints | Auto label with your own weights |
| Classification | Image labels | Manual labeling |
See Annotation types and tools for how each tool behaves.
How BOAT is put together
BOAT is made of a few small pieces that you rarely need to think about:
- The
boatCLI is a single Go binary. It installs, configures, starts, updates and backs up everything else. - Core is the web application: a Python backend and the annotation editor, served on
http://127.0.0.1:7234by default. - Model packages are installed separately. Each one runs in its own isolated Python environment, called a worker, so models never fight over dependencies.
- The model registry at
models.boat-annotation.comhosts signed package manifests, code archives and weights. The Models page on this site is a snapshot of it.
Supported platforms
| Platform | Architecture | Accelerators |
|---|---|---|
| Ubuntu 22.04 / 24.04 | x86_64 | NVIDIA CUDA 12.6 or 12.8 (Blackwell), CPU |
| macOS | Apple silicon (arm64) | Metal (MPS), CPU |
The annotation service itself never needs a GPU. Models run on whatever you have, and each model page lists the runtime profiles it supports.
Next steps
- Quickstart: install, start and label your first image in a few minutes.
- Installation: what the installer does, file locations and uninstalling.
- CLI reference: every
boatcommand and flag. - Model library: compare models and copy an install command.