Most vision integrations demo the good case and let you discover the failure modes yourself. Putting the actual frame beside what the model said about it tells you what resolution and framing you will need to get a useful answer. The obvious caution is the obvious one: this connects a camera in your room to a model, so it belongs on a machine and in a moment where you have decided that is fine.
A small Python server that gives an assistant access to a connected camera through OpenCV, returning stills it can then examine.
Access to and control of a connected webcam, with captured stills handed back for the model to analyse. Their README shows the captures beside what the model made of them.
Python 3.10+, OpenCV and the MCP Python SDK; uv optional. MIT licensed.
Build from source — clone the repository and build it, then point your client at the binary