Images

One of the most effective ways to give Pi useful context is to share an image. A picture really is worth a thousand words — instead of describing a design you like or a bug you’re seeing, you can just show it.

This only works if the model you’re using has vision capability. Vision models can look at an image and understand what’s in it: they can describe photographs, extract text from screenshots, and interpret UI details like layout, spacing, typography, and color.

How to add an image

Pi supports a few ways to bring an image into a conversation:

• Paste — copy an image to your clipboard and press Ctrl+V (Alt+V on Windows).
• Drag and drop — drag an image file into a terminal that supports it.
• @ file reference — type @ and fuzzy-search for a file by name, including images.
• CLI argument — attach a file directly when starting a one-off prompt: pi @screenshot.png "What's in this image?"

Whichever method you use, the image becomes part of your message just like text — you can combine an image and a prompt in the same message.

Image settings

Two settings control how Pi handles images, in ~/.pi/agent/settings.json or a project’s .pi/settings.json:

{
  "images": {
    "autoResize": true,
    "blockImages": false
  }
}

Which models support vision

Not all models can interpret images, and vision support isn’t tied to a specific subscription product the way it might be with other tools — it depends on which model you’re actually using:

• OpenAI’s Codex model (the /login ChatGPT Plus/Pro path from Installation) supports image input.
• Free OpenRouter models — some :free-tagged models support vision too, though the specific list changes over time as providers add and remove models. Check openrouter.ai/models and filter for models tagged :free with image input support.

If a model doesn’t acknowledge your image or seems to ignore it, that’s a sign to switch to a confirmed vision-capable model.

What vision is useful for

Here are some concrete ways images help:

Replicate a design — share a screenshot of a UI you admire and ask Pi to recreate it. The model can read the layout, colors, spacing, and typography from the image.

Debug a visual bug — share a screenshot of something that looks wrong. The model can see what you’re seeing and help diagnose it without you having to describe every detail.

Implement a mockup — share a wireframe or design file export and ask Pi to build it. This is often faster and more accurate than trying to describe the design in words.

Extract text from an image — share a screenshot containing text (a menu, a dialog, an error message) and the model can read it for you.

Try it

Make sure you’re using a vision-capable model, then try one of these:

• Take a screenshot of any webpage or app and ask: What do you see in this image?
• Share a screenshot of a UI and ask: How would you describe the design of this interface?

If the model responds with details from the image, vision is working.