Skip to main content

Claude Code Integration

CROAK integrates natively with Claude Code through slash commands. This is the recommended way to use CROAK for an interactive, guided experience.

Setup​

  1. Initialize your project — this automatically sets up Claude Code integration:

    npx croak-cv init

  2. Open your project in Claude Code (VS Code with Claude extension, or Claude Code CLI)

  3. Start with the Router — type /croak-router to get guidance on next steps

When you run croak init, CROAK creates:

  • .claude/skills/croak-*/SKILL.md — Skill files for each agent
  • CLAUDE.md — Project context file that Claude Code reads automatically

Claude Code discovers these files and makes them available as slash commands.

Agent Commands​

Each command activates a specialized AI persona with domain expertise, guardrails, and a knowledge base.

CommandAgentWhat It Does
/croak-routerDispatcherStart here. Pipeline coordinator that guides you through the workflow
/croak-dataScoutScan directories, validate images, manage vfrog SSAT or classic annotations
/croak-trainingCoachConfigure training across local GPU, Modal, or vfrog platform
/croak-evaluationJudgeEvaluate models, analyze errors, generate reports
/croak-deploymentShipperDeploy to vfrog inference, Modal serverless, or edge devices

Workflow Commands​

End-to-end pipelines that chain multiple agent steps together.

CommandDescription
/croak-data-preparationFull data pipeline: scan, validate, annotate, split, export
/croak-model-trainingTraining pipeline: recommend, configure, execute, handoff
/croak-model-evaluationEvaluation pipeline: evaluate, analyze, diagnose, report
/croak-model-deploymentDeployment pipeline: export, optimize, deploy, verify

Example Session​

You: /croak-router

Claude: Dispatcher here! I see this is a new CROAK project.
        Current stage: uninitialized

        Let me help you get started. Do you have images ready to train on?

You: Yes, I have 500 product images in ~/photos/products

Claude: Great! Let me hand you off to Scout (Data Agent) to scan
        and validate them.

You: /croak-data

Claude: Scout reporting for duty! I'll help you prepare your dataset.
        Let me scan ~/photos/products...
        [Runs: croak scan ~/photos/products]

        Found 500 images. 487 valid, 13 have issues...

Agent Details​

Dispatcher (Router)​

The pipeline coordinator. Tracks your progress through the workflow stages, recommends next actions, and routes you to the right specialist agent.

Commands: status, init, reset, next, help, vfrog setup

Scout (Data)​

Data quality specialist. Scans directories for images, validates annotations, checks class balance, and manages the annotation workflow — either vfrog SSAT or classic import.

Commands: scan, validate, convert, split, annotate, stats, visualize

Quality guardrails:

  • Minimum 100 images, 50+ per class
  • Maximum 10:1 class imbalance
  • Annotation coverage checks
  • Corrupt image detection

Coach (Training)​

Training specialist. Recommends model architectures based on your dataset, estimates training cost and time, configures hyperparameters, and manages experiment tracking via MLflow or Weights & Biases.

Supported architectures: YOLOv8, YOLOv11, RT-DETR

Supported providers: Local GPU, Modal.com, vfrog platform

Judge (Evaluation)​

Evaluation specialist. Calculates metrics (mAP, precision, recall, F1), performs error analysis, identifies failure patterns, and generates detailed performance reports.

Metrics: mAP@50, mAP@50-95, per-class precision/recall, confusion matrices

Shipper (Deployment)​

Deployment specialist. Exports models to optimized formats, handles quantization, and deploys to your target environment.

Export formats: ONNX, TensorRT, CoreML, TFLite, TorchScript

Deploy targets: vfrog inference API, Modal serverless, edge devices

Agent Handoffs​

Agents communicate through validated handoff contracts. When one agent completes its work, it produces a structured artifact that the next agent can pick up. For example:

  1. Scout validates the dataset and produces a DatasetArtifact
  2. Coach receives the artifact, trains, and produces a ModelArtifact
  3. Judge evaluates the model and produces an evaluation report
  4. Shipper deploys using the model and evaluation results

Handoff files are stored in .croak/handoffs/ and can be inspected for debugging.

Next Steps​

Last updated on