> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.ovalbee.com/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.ovalbee.com/_mcp/server.

# Welcome to OvalBee 🐝

OvalBee is a platform for computer vision teams. You describe the whole process as a **flow**: a graph of nodes on a shared canvas that covers where images come from, how they get labeled and reviewed, which model trains on them and how it's evaluated. The graph isn't a diagram of work happening somewhere else. It's the system the team runs, watches and changes.

![An OvalBee flow that profiles a dataset, predicts with Grounding DINO and runs quality, embedding and label diagnostics, each with its own report](/_fern-files/_img/e0e7e7719277529a211d513c05b5d2808d59c199484e09588803b30a786fa42f.webp)

## Built for computer vision

OvalBee isn't a general workflow tool with a few vision plugins. Everything in it speaks the language of CV work:

* **Nodes for the real tasks.** Import and version datasets, sample and split, label and review, train and predict with YOLO, DEIM or EdgeCrafter, evaluate and compare detectors, find duplicates, outliers and label errors with embeddings. You connect ready steps instead of writing glue code.
* **The entities you already think in.** Datasets with versions, annotations, ontologies with versioned classes and tags, labeling queues with per-person batches, a model registry with lineage and benchmarks.
* **Reports that answer the question.** Evaluation, training and data quality nodes each build a report that states its verdict first and backs it with dozens of charts, per-class tables and image galleries you can drill into. A full, reproducible model evaluation takes three nodes and one run.

A flow keeps working after the first run. New images land in a pool, labelers pull batches, rejected work goes back for another pass, and a training node picks up the latest dataset version when you run it again. Every run, dataset version, model and report stays traceable to what produced it.

## Three ways to build a flow

#### [Start from a template](/build/templates)

Pick a ready workflow for importing, labeling, training or evaluation, then change any step.

#### [Ask an AI agent](/build/agent)

Describe what you need in the built-in AI Chat, or connect Claude Code, Codex or another MCP client.

#### [Build on the canvas](/build/editor)

Add nodes from the Node Library, connect them and bind their inputs yourself.

All three produce the same thing: a regular flow you can open, edit and run. Mixing them is normal: start from a template, ask the agent to add an evaluation branch, then fine-tune thresholds by hand.

## Start here

#### [Core ideas](/start/core-ideas)

The five ideas that make the rest of the docs easy to read. Five minutes.

#### [Quickstart](/start/quickstart)

Import a public dataset and get a quality report on it. About ten minutes.

#### [Use cases](/use-cases/overview)

Labeling with review, training, evaluation and data cleanup, each with a ready template.

#### [Node library](/node-library/nodes/overview)

Every node you can place on the canvas, with its inputs, outputs and behavior.

## Concepts

#### [Flows](/concepts/flows)

#### [Nodes](/concepts/nodes)

#### [Edges](/concepts/edges)

#### [References](/concepts/references)

#### [Runs and automation](/concepts/runs)

#### [Datasets and versions](/concepts/datasets)

#### [Labeling](/concepts/labeling)

## Running OvalBee for a team

Workspaces, roles, secrets, compute and sign-in are covered in the [Administration](/administration/overview) section. Start there if you're setting OvalBee up for others rather than building flows yourself.

> **Note**
>
> OvalBee is in private beta. The product changes weekly and so do these docs. If something here doesn't match what you see in the app, tell your OvalBee contact. We'd rather fix the docs than have you guess.