#Image Attachments

When you paste or autocomplete an image path, crabcode inserts a placeholder such as [Image #1]. The placeholder is rendered as interactive text: hovering changes the text color only, and clicking it opens the image.

By default, image opening uses automatic detection:

{
  "images": {
    "openWith": "auto",
  },
}

#Open target

images.openWith controls where clicked image placeholders open.

auto treats standalone terminals such as WezTerm, iTerm, Terminal.app, Alacritty, and Kitty as normal terminals, so they open images with the OS-level file opener.

#Custom command

Use a command object when you want complete control over the opener:

{
  "images": {
    "openWith": {
      "command": "zed",
      "args": ["{path}"],
    },
  },
}

Other examples:

{
  "images": {
    "openWith": {
      "command": "open",
      "args": ["-a", "Preview", "{path}"],
    },
  },
}

If args is omitted, crabcode uses ["{path}"].

#Vision subagent

Some coding models can work with images only when another vision-capable model describes the image first. Crabcode ships a built-in vlm-agent subagent for that workflow instead of sending raw images to models that reject image input.

Configure the built-in subagent with a vision-capable model:

{
  "agent": {
    "vlm-agent": {
      "model": "xai/grok-4.3",
    },
  },
}

When the active model cannot receive raw images and agent.vlm-agent.model is configured, Crabcode exposes vlm-agent as an available subagent and includes the existing local image paths in the text prompt. The active model is instructed to call the task tool with subagent_type: "vlm-agent"; the subagent uses view_image directly on those paths and returns visual findings for the active model to use in its answer. During this fallback turn, the parent model is not given direct view_image access, so image inspection goes through vlm-agent instead of the parent guessing a path.

If the active model cannot receive raw images and agent.vlm-agent.model is not configured, Crabcode shows a warning toast with the config hint and sends a text-only note instead of making a failing image request.