LLM gave wrong json input for mcp tool

[liHai001]

What happened?

call mcp tool to get chart,call llm got wrong json format input to call tool. I can success using Cherry studio or DeepChart with same LLM and MCP server. Where can i change system prompt for tool calling or change something else?

Version Information

ghcr.io/danny-avila/librechat-dev latest
ghcr.io/danny-avila/librechat-rag-api-dev-lite latest
ghcr.io/librechat-ai/codesandbox-client/bundler latest
ghcr.io/virtuos/librechat_exporter main

Steps to Reproduce

1. Local LLM service: Qwen3-30B
2. MCP: https://github.com/antvis/mcp-server-chart
3. chat text: "Draw a line chart: Sales of 10 million in 2020, 20 million in 2021, 5 million in 2022, and 15 million in 2023"
4. call mcp tools correct, but got wrong json format as tool's input: {"data": "2020", "value": 10}, {"time": "2021", "value": 20}, {"time": "2022", "value": 5}, {"time": "2023", "value": 15}]}

What browsers are you seeing the problem on?

Microsoft Edge

Relevant log output

no error logs on backend

Screenshots

request again:

Code of Conduct

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

[danny-avila]

Using the same model with openrouter is working:

Same with similar model, using NVIDIA AI gateway:

[liHai001]

I serving local Qwen3-30B using VLLM, maybe that's the difference?
I'm not sure, when I use same model service(by vllm),same mcp tool, DeepChat desktop can generate json format for mcp tool input correctly. But librechat generate wrong json format.

my endpoint:

custom:
- name: "Qwen3-30B"
apiKey: "vllm"
baseURL: "http://myserverip:port/v1"
models:
default: ['Qwen3-30B-A3B']
fetch: true
titleConvo: false
titleModel: "Qwen3-30B-A3B"
titleMethod: "completion"
titlePrompt: "Create a brief, descriptive title for the following conversation:\n\n{convo}\n\nTitle:"
summarize: false
summaryModel: "current_model"
forcePrompt: false