Inspect an image with a vision tool

ViewImage sends a local image to a vision model and returns an observation. Use it when a task needs something read from a screenshot, a diagram, or a photo.

Before you start

Steps

  1. Ask about a specific file, and say what to look for:

    Open docs/screenshots/error.png and tell me which field is highlighted.
    
  2. The agent calls ViewImage with the file path and a prompt. You get back a visual observation plus metadata such as media type, byte size, and dimensions.

  3. If you ask about the same image and prompt again, the runtime reuses the cached observation instead of calling the model again.

Confirm it worked

The observation answers the question you asked, and the metadata confirms the tool read the file you meant.

Limits

Supported formats, model selection, and the full response shape are in the Image observation reference.