Try It Out — OpenEye Claude Skill

The OpenEye Claude Skill lets you use Claude (via Claude Code or the Claude Desktop app with MCP) as a copilot for your vision AI workflows. Ask Claude to pull models, run inference, manage fleets, and more — all through natural language.

What is a Claude Skill?

A Claude Skill is a structured knowledge file that teaches Claude about a specific tool or API. When loaded, Claude can understand OpenEye's CLI commands, API endpoints, model registry, and architecture — letting it write correct commands and code on your behalf.

Prerequisites

• Claude Code (CLI), or Claude Desktop / another MCP client for the MCP server
• OpenEye CLI installed (pipx install openeye-sh)
• A terminal or development environment

Setup — Claude Code

Step 1: Clone the OpenEye repo

git clone https://github.com/aw3-technology/openeye.sh.git
cd openeye.sh

Step 2: Start Claude Code

claude

Claude Code automatically discovers the .claude/ directory in the project root, which contains the OpenEye skill files. No extra configuration needed.

Step 3: Verify it works

Ask Claude:

What OpenEye models are available and how do I pull one?

Claude will reference the skill files and give you accurate, project-specific answers.

Setup — MCP Server (Claude Desktop and other MCP clients)

The CLI ships an MCP server, openeye-mcp, that lets Claude call OpenEye directly: list, pull, and remove models, run inference on an image, start a server, check its health, benchmark, and read Nebius stats. It runs over stdio on your machine.

Step 1: Install with the mcp extra

pipx install "openeye-sh[mcp]"

Step 2: Register the server

Claude Code:

claude mcp add openeye -- openeye-mcp

Claude Desktop — add it to claude_desktop_config.json (Settings → Developer → Edit Config), then restart the app:

{
  "mcpServers": {
    "openeye": { "command": "openeye-mcp" }
  }
}

If Claude Desktop can't find the command, use the full path from which openeye-mcp.

Step 3: Ask for something

List the OpenEye models and run yolov8 on ~/Pictures/desk.jpg

Example Prompts

Here are things you can ask Claude with the OpenEye skill loaded:

Model Management

Pull the YOLOv8 model and run inference on photo.jpg

What quantized model variants are available?

Benchmark the depth-anything model with 50 runs

Server & API

Start serving YOLOv8 on port 8000 and show me how to call the predict endpoint with curl

Write a Python script that connects to the OpenEye WebSocket and streams camera frames

Fleet Management

Register a new camera device called "lobby-cam" and deploy YOLOv8 to it with a canary strategy

Show me all online devices in my fleet

MLOps

Upload my custom ONNX model and set up an A/B test between v1 and v2

Export the YOLOv8 model to TensorRT with quantization

Architecture & Debugging

Explain the perception pipeline and how plugins are loaded

I'm getting "Connection refused :8001" — help me debug this

Skill File Structure

The skill is organized under .claude/ in the OpenEye repository:

.claude/
  skills/
    openeye/
      SKILL.md              # Skill entry point: overview & quick reference
      CLI_REFERENCE.md      # All CLI commands with options
      API_REFERENCE.md      # Server endpoints, TypeScript clients, hooks
      ARCHITECTURE.md       # Backend pipeline, mode system, plugins
    openeye-run/
      SKILL.md              # /openeye-run: execute OpenEye CLI workflows

Tips for Best Results

• Be specific — Instead of "run a model," say "run YOLOv8 on image.png with pretty output"
• Chain tasks — "Pull depth-anything, benchmark it, then start serving it on port 9000"
• Ask for code — "Write a JavaScript client that calls the /predict endpoint and displays bounding boxes on a canvas"
• Debug with context — Paste error messages and ask Claude to diagnose using its knowledge of OpenEye's error codes and architecture
• Reference the docs — "According to the OpenEye API reference, what query parameters does /predict accept?"

What Claude Can Help With