Open-source NVIDIA Vision MCP for ZCode / GLM

[cyril-belin]

NVIDIA Vision MCP for ZCode / GLM

Hi ZCode team,

I’ve been testing ZCode/GLM on a real Flutter UI reproduction workflow using reference screenshots.

One limitation I ran into is that the coding model itself does not always have native image understanding available in the workflow. I initially tested Gemini Vision through MCP, but eventually hit quota limits.

So I built a small open-source MCP server that gives ZCode access to vision models through NVIDIA NIM:

https://github.com/cyril-belin/nvidia-vision-mcp

What it does

It exposes a simple MCP tool:

analyze_image

with support for:

• image_path for local stdio usage
• image_url
• image_base64
• configurable vision model
• local stdio transport
• remote Streamable HTTP transport

The current default model is:

minimaxai/minimax-m3

through NVIDIA NIM.

Why I built it

The interesting part for me was not only making vision available, but seeing how ZCode actually used it during an autonomous coding loop.

In my Flutter test, ZCode was able to:

1. inspect the reference UI
2. implement the screen
3. launch the iOS simulator
4. take screenshots
5. call the vision MCP
6. compare the implementation against the reference
7. correct visual differences
8. repeat the loop autonomously

I did not need to manually prompt it after every visual correction.

That made MCP-based vision a surprisingly effective workaround for models/workflows without native image input.

Rate limiting

During testing, NVIDIA free endpoints occasionally returned HTTP 429.

The MCP therefore includes:

• serialized vision requests
• automatic 429 retry
• Retry-After handling
• exponential backoff
• protection against long quota waits

This was tested against real NVIDIA throttling.

Open source

The repository is MIT licensed and designed so every user supplies their own NVIDIA API key.

It can run locally with ZCode or be self-hosted remotely.

Repository:

https://github.com/cyril-belin/nvidia-vision-mcp

I’m sharing it here mainly because it may be useful to other ZCode users working with UI screenshots, visual QA, or image-based coding workflows.

I’d also be interested to know whether native image/vision support is planned directly inside ZCode/GLM workflows in the future.