Moondream Local Vision

Problem

When engineers need to inspect local images in chat or automation without sending raw pixels to an external vision API, moondream-vision-zc provides a lightweight path. It connects a local Moondream model through Ollama to OpenClaw, so an image or local file path can produce a natural-language description.

How It Works

• The skill depends on a local Ollama service and the Moondream model, sending base64 image data and optional prompts via the Ollama REST API.
• After registration in OpenClaw, it can be triggered with a command such as !moondream <image_path>.
• The returned description can be passed to GPT-OSS-120B for summarization, analysis, drafting, or cross-modal reasoning.
• Moondream is positioned as a lightweight local inference model, with the source noting roughly 1-2 seconds per image on a typical laptop CPU and further speedup on GPU.

Boundaries

• A local Ollama service must be reachable, including port 11434.
• Ollama has roughly a 5 MB per-image limit, so large images should be compressed first.
• If output is empty, check that the prompt is non-empty or include an empty system field in the payload.
• It is best for quick local image description and upstream understanding; complex conclusions still benefit from a second pass with a larger language model.