Vision
Vision lets nod inspect screenshots, diagrams, and other visual context during a task.
How vision works
Every image goes to the session model; nod has no separate image helper.
Native vision
Images you attach with /image, --image, or an ACP image block are included in the same request for the model to read directly. Both Codex and Grok models accept images this way.
The vision tool
The model can also call the built-in vision tool when it wants to look at an image on its own initiative: one of the attached images by ID, or a file path in the workspace. The tool sends those images with a focus question to the session model in a separate request and returns the description as text, bounded by image_adapter_output_bytes. This keeps large images out of the main conversation until they are needed.
It is a separate model request with its own token usage; see Additional model requests.
Add visual context
Interactive shell
Attach a file:
/image ./diagram.png
/img is an alias. You can also type an image path directly in a prompt.
Use /images to inspect pending attachments or /images clear to remove them.
On macOS, attach an image from the clipboard:
/paste
Headless requests
Attach an image to a one-off request:
nod ask --image ./ui.png "describe this interface"
Repeat --image to attach more than one file.
Supported inputs
nod accepts PNG, JPEG, GIF, and WebP images up to 10 MiB each.
Permissions and trust
The vision tool follows the active permission mode. It reads only images attached to the session or image paths within the workspace scope. Instructions found inside images are treated as untrusted content.