delivery-vision-mcp

An MCP (Model Context Protocol) server that exposes trained computer vision models as tools an AI agent can call directly, built for logistics and delivery use cases like Pathao, Redx, Carrybee, Steadfast etc.

Instead of connecting to someone else's vision service, this server wraps your own trained YOLO models so any MCP client (Claude Desktop, Claude Code, Cursor) can call them mid-conversation.

Tools

Related MCP server: mcp-see

What makes this more than a demo

• Severity is more than confidence. detect_package_damage blends detection confidence with how much of the frame the damage covers, so a small high-confidence scuff doesn't register the same as a crushed side.

• Batch support. detect_damage_batch checks a whole set of images (e.g. one hub's daily intake) in a single tool call instead of one at a time.

• No raw crashes. Every tool validates the image path and model weights first and returns a clean, readable message on failure, missing file, corrupt image, missing weights, instead of a stack trace reaching the agent.

• Self-reporting. list_available_tools shows which models are actually loaded, so an agent (or you) can check readiness before calling a tool that will fail.

⚠️ Attention: This Won't Work Out of the Box

Before you get started, it's worth being upfront about one thing: this repo gives you the plumbing, not the intelligence. The server, the tool definitions, the error handling, the severity scoring, all of that is done. What it doesn't come with are the actual computer vision models.

That part is on you. You'll need to train your own YOLO model for each capability you want to use, whether that's spotting a dented parcel, counting boxes in a hub photo, checking if a rider has their helmet on, or telling a CNG apart from a rickshaw. Once you've trained a model and have a .pt weights file in hand, you just drop it into the models/ folder with the right name, and that tool switches on automatically. No code to touch, no config to rewrite.

Until you do that, the server will still run and Claude will still see all six tools, but calling one without weights behind it just gets you a polite "not ready yet" message instead of a real answer. So think of this less as a finished product and more as a well-built frame waiting for you to fill in the glass. The actual computer vision work, collecting images, labeling them, training the model, is the part that turns this from a nice piece of scaffolding into something that genuinely sees.

Setup

1. Install dependencies:

   pip install -r requirements.txt

2. Add trained weights to models/, one file per tool you want live:

   • models/parcel-damage.pt — damage type detection

   • models/parcel-count.pt — box/parcel counting

   • models/helmet-detection.pt — helmet vs no-helmet detection

   • models/vehicle-classify.pt — vehicle type classification

   You don't need all four at once. Any tool whose weights are missing returns a clear setup message instead of crashing the server, so you can build this out incrementally.

3. Run the server locally to confirm it starts:

   python server.py

Connect to Claude Desktop

Add this to your claude_desktop_config.json:

{
  "mcpServers": {
    "delivery-vision": {
      "command": "python",
      "args": ["/full/path/to/delivery-vision-mcp/server.py"]
    }
  }
}

Restart Claude Desktop. It will list all six tools automatically and can call them when you share an image path (or a list of paths, for the batch tool) in conversation.

Why build this instead of just using an existing MCP server

Anyone can connect to Roboflow's or Groundlight's MCP server. This project demonstrates the other half of the skill: training your own CV models and exposing them as agent-callable tools, the part that actually requires computer vision expertise.