BOAT
Download
FeaturedSAM3 LiteText scores 10/10 in the model library

Label images with local AI. Keep your data on your machine.

BOAT is a free, open-source annotation tool that runs on your own machine. Install it and connect models with a single command.

curl -fsSL https://boat-annotation.com/install.sh | sh

macOS on Apple silicon Ubuntu 22.04 / 24.04 x86_64

DownloadBrowse models
annotation types
7annotation types
open models
7open models
AI tools
22AI tools
cloud dependencies
0cloud dependencies

Works with open models

  • SAM3 LiteText
  • SAM3
  • PET-DINO
  • LocateAnything
  • S3OD
  • LLMDet
  • Falcon Perception
  • YOLO (bring your own weights)

One tool, seven tasks

Every annotation type,
with AI where it helps

Boxes, rotated boxes, masks, pixel classes, keypoints, points and image labels, all in one editor. Prompt a model, review its suggestions and keep what's right.

127.0.0.1:7234Object Detection / highway_0042.jpg
Object Detection example
SAM3 LiteText↵

Tool

Text to Boxes

Labels

  • Car3
  • Truck1
  • Bus1
Saved⌘S

Up and running in a minute

Three commands from zero to auto-labeling

  1. Install the CLI

    One line. The installer checks your platform, verifies the checksum and Sigstore signature, and puts boat in ~/.local/bin.

  2. Start BOAT

    Runs on 127.0.0.1:7234 with SQLite and local files. No Docker, database server or account with us.

  3. Add a model

    Each model gets an isolated runtime and verified weights. Its tools show up in the editor within seconds, without a restart.

Model library

Pick a model.
Copy one line.

Every model is a signed package with its own tools, runtimes and memory needs. Compare them, then install the one that fits your hardware.

Browse all 7 models

SAM3 LiteText

sam3-litetext · EfficientSAM3

10

SAM3 with a distilled MobileCLIP text encoder. Keeps SAM3's text, box and point prompts while using less memory — the best default for most machines.

DetectionInstance seg.Semantic seg.Oriented boxes
2.2–2.5 GBmacOSCUDACPU

SAM3

sam3 · Meta

8

Meta's Segment Anything Model 3. Promptable concept segmentation from text, boxes or points — the most complete toolset in the library.

DetectionInstance seg.Semantic seg.Oriented boxes
3.5 GBmacOSCUDACPU

PET-DINO

pet-dino

7

Visual-prompt detector: draw one box and it finds every similar object. Pairs with Cascading for fast, repetitive detection work.

Detection
2.8 GBmacOSCPU

LocateAnything

locate-anything · NVIDIA

6

NVIDIA's 3B grounding model, quantized to GGUF and run with llama.cpp on Apple silicon. Text to boxes, or text to points for counting crowds.

DetectionCounting
3–7.7 GBmacOS

S3OD

s3od

6

Lightweight salient-object segmentation. Draw a loose box and it snaps to the object inside — as a mask or a rotated box — in under 1 GB of memory.

Instance seg.Semantic seg.Oriented boxes
0.5 GBmacOSCUDACPU

LLMDet

llmdet · iSEE Lab

4

Open-vocabulary detector trained with large-language-model supervision. Type what you are looking for and get bounding boxes.

Detection
1.4–5 GBmacOSCPU

Why BOAT

Serious labeling without the infrastructure

Everything runs on one machine you control, and the parts that need to be fast, private and verifiable are.

Your data stays home

BOAT binds to 127.0.0.1 by default and stores projects in a local SQLite database and your filesystem. Images never pass through someone else's server.

Nothing else to run

No containers, database servers, email service or paid plan. One binary and a data folder.

Models plug in live

Each model installs into its own isolated runtime. New tools appear in the editor within seconds, with no restart.

Runs on the hardware you have

Apple silicon with Metal, NVIDIA GPUs with CUDA 12.6 or 12.8 (Blackwell included), or plain CPU. Set a device per model; BOAT reserves memory before loading and frees idle workers when another model needs room.

Built for small teams

A personal workspace for everyone and unlimited team workspaces, with Owner, Admin, Labeler and Viewer roles.

Verified end to end

Sigstore-signed releases, an Ed25519-signed model index and SHA-256-pinned weights.

Backups and rollbacks

Coordinated backups, checked updates and a one-command rollback if an update misbehaves.

Start labeling
in under a minute

One command installs BOAT. A second starts it. Your first model is a third.

curl -fsSL https://boat-annotation.com/install.sh | sh
Read the quickstart Other install options
↑↓ navigate↵ open