[Bug]: MCP Image Generation Tools: Generated images sent to non-vision models causing errors

[JumpLink]

What happened?

When MCP (Model Context Protocol) tools return ImageContent (e.g., from image generation tools), LibreChat correctly converts them to artifacts for display. However, these artifacts are then sent back to the LLM in subsequent requests, which causes errors for non-vision models.

Expected behavior:
Image generation tools should work like LibreChat's built-in image generation tools (imageGenTools set): the generated images should be displayed to the user but NOT sent back to the LLM.

Actual behavior:
After an MCP tool generates an image successfully, LibreChat attempts to send the image artifact to the LLM, resulting in errors like:

• "404 No endpoints found that support image input"
• "MODEL_NOT_FOUND"

This happens because non-vision models (e.g., Deepseek) don't support image inputs.

Version Information

ghcr.io/danny-avila/librechat:v0.8.2-rc3

Steps to Reproduce

1. Set up an MCP server that returns ImageContent (e.g., image generation tool)
2. Configure the MCP server in LibreChat's librechat.yaml
3. Use a non-vision model (e.g., Deepseek, GPT-3.5)
4. Call the MCP image generation tool in a conversation
5. The image is generated and displayed successfully
6. LibreChat then attempts to send the image artifact to the LLM
7. Error occurs: "404 No endpoints found that support image input" or "MODEL_NOT_FOUND"

Example MCP Tool Response:

{
  content: [
    { type: 'text', text: 'Image generated successfully.' },
    { type: 'image', data: 'iVBORw0KGgo...', mimeType: 'image/png' }
  ]
}

What browsers are you seeing the problem on?

No response

Relevant log output

2026-01-20 10:14:45 error: [api/server/controllers/agents/client.js #sendCompletion] Operation aborted 404 No endpoints found that support image input
Troubleshooting URL: https://js.langchain.com/docs/troubleshooting/errors/MODEL_NOT_FOUND/⁠

Screenshots

Code of Conduct

• I agree to follow this project's Code of Conduct

[danny-avila]

Image generation tools should work like LibreChat's built-in image generation tools (imageGenTools set): the generated images should be displayed to the user but NOT sent back to the LLM.

Image-gen tools do send the image back to the LLM on a provider basis but the MCP output may not be making the exception. In my opinion, image outputs are better if they are sent back to the LLM if they are vision capable. I’ll look into this further.

[JumpLink]

Image-gen tools do send the image back to the LLM on a provider basis but the MCP output may not be making the exception. In my opinion, image outputs are better if they are sent back to the LLM if they are vision capable. I’ll look into this further.

Yes, that's right. In my case, the MCP generates the images, so it makes less sense, but there are other MCP tools where it would make much more sense to send them back to the model, for example an MCP for image search, where it would make perfect sense for the model to also recognise the image.

[JumpLink]

In order for vision to be taken into account, support for vision?: boolean would probably need to be added to modelSpecs, as I suggested in #11333

[danny-avila]

Closing in favor of #11418