llama-server fails when MCP tool returns images.

[Sultan-papagani]

I’m using llama-server with an MCP tool that generates images. When the LLM sends a message containing an image generated by the tool, I get the following error when i reply to the LLM:

image input is not supported - hint: if this is unexpected, you may need to provide the mmproj

I understand that llama-server (with a text-only model) cannot handle image inputs, but in my use case I don’t want to send MCP-generated images back to the LLM at all.

What should i do ?

Thanks!

[CAHbKA-IV]

You need to use multimodal model. It has two files: model file amd multimedia projector file. For example, at https://huggingface.co/unsloth/Qwen3.5-9B-GGUF/tree/main you can find such model. To load it you must use --mmproj command line parameter:
./llama-server --model ./Qwen3.5-9B-BF16.gguf --mmproj ./mmproj-F16.gguf
Aware, .mmproj files are different for each model, files from other models will cause errors!

[Sultan-papagani]

Thanks, but as I explicitly mentioned in the second paragraph: I already know it's a text-only model and I do not want to send the images back to the LLM.