How to connect LLM via OpenAI compatible API

[murarduino]

When I used langchain and Langgraph projects before, I was used to connecting to local LLM services using openAI compatible APIs, so I didn't need to consider the troubles caused by different LLM inference engines (llama.cpp/vllm/MindIE can access local LLM running through OpenAI compatible APIs).
So I hope open_deep_research can support this feature as well. Can I do this? How should we do it?

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(openai_api_base="http://10.31.13.16:8088/v1",api_key="123321", model="Qwen3-14B")

return_object=llm.invoke("你好，请介绍一下自己")
print(return_object.content)

[Rainier-rq]

simply use ollama

[yanganhui]

I’d like to clarify the solution a bit more precisely.

The model selection itself (model names and MAX_TOKENS) was already supported by open_deep_research. No changes are needed there.

The only missing piece was the ability to configure a custom LLM_BASE_URL, so that Cursor can inject an OpenAI-compatible base URL (for example, a locally deployed LLM server or DeepSeek endpoint).

Once LLM_BASE_URL is configurable, everything works out of the box, because the rest of the model configuration is already handled via environment variables.

Example .env:

LLM_BASE_URL=https://api.deepseek.com/v1

DEEPSEEK_API_KEY=sk-xxxxx

SUMMARIZATION_MODEL=deepseek:deepseek-chat
SUMMARIZATION_MODEL_MAX_TOKENS=8192

RESEARCH_MODEL=deepseek:deepseek-chat
RESEARCH_MODEL_MAX_TOKENS=8192

COMPRESSION_MODEL=deepseek:deepseek-chat
COMPRESSION_MODEL_MAX_TOKENS=8192

FINAL_REPORT_MODEL=deepseek:deepseek-chat
FINAL_REPORT_MODEL_MAX_TOKENS=8192

So in short: model configuration already works; adding support for LLM_BASE_URL just enables Cursor to route requests to any OpenAI-compatible backend.