Moondream is small enough to run on a laptop, and that changes what you are willing to point it at. Screenshots, scanned paperwork and camera frames get answered locally instead of being uploaded to somebody's API.
A FastMCP server for Moondream, a small vision-language model. The model runs locally, and the server exposes image captioning, visual question answering, object detection and visual pointing.
- Short, normal or detailed captions for an image
- Natural-language questions answered about an image
- Objects detected and located with bounding boxes
- Precise coordinates returned for an object in an image
- Local files and remote URLs both accepted
- Batch processing across multiple images
- Device detected and optimised for automatically across CPU, CUDA and Apple Silicon MPS
Python 3.10 or higher and a recent PyTorch with support for your device. uvx moondream-mcp runs it without installing anything. MOONDREAM_DEVICE controls device selection and defaults to auto.
One command — uvx moondream-mcp
