Skip to content

The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.

2.2 KB
Raw Blame

description: >- Show the agent an image: attachments the model can see, local image metadata, and which model handles vision. icon: image

Image tools

This page is about getting an image into a turn so the model can look at it. To make images instead, see Image and video generation.

Attaching an image

Attach a file in the composer, or point the agent at a path. What happens next depends on the model and the route, not on a setting.

  • If the model accepts images and the transport supports them, the image goes up natively as image content.
  • If not, it degrades instead of failing. The agent gets a bounded readout, or a reference to the file with its path and metadata. It can still reason about the file and tell you what it cannot see.

Uploads stay in the acting workspace across a restart. The ordered references stay in the transcript, so a later turn can still refer to "the second screenshot".

image_info

image_info is a read-only tool. It reads a local image's metadata (format, dimensions, size) and can return the bytes as base64 text for a model that wants them inline. Use it when the question is "what is this file". It does not need a vision model.

Which model sees it

Vision has its own provider slot, so you can route it separately from chat:

vision_provider = "…"   # a provider id, or leave unset to use the chat provider

If a configured vision model turns out to be chat-only, the call returns a named error. It never silently swaps in another model. There is no default local vision model, so a local setup has to name one. Local models and BYOK lists the ones worth trying.

Delegation

When a turn needs to look at several images, or study one closely, the agent usually hands it to the built-in vision sub-agent. That keeps the main conversation from carrying every pixel of context. The agent decides this itself. You do not have to ask.

See also