LangGraph with a custom OpenAI-compatible model

1. 01

   Bind the model

   Create a ChatOpenAI client pointed at the gateway and bind the tools each node needs.

   python

   from langchain_openai import ChatOpenAI
   from langgraph.prebuilt import create_react_agent
   
   model = ChatOpenAI(
       model="gpt-6-sol",
       base_url="https://uncensoredai.dev/v1",
       api_key="sk-ua-XXXX-XXXX-XXXX-XXXX",
   )
   agent = create_react_agent(model, tools=[search, calculator])

2. 02

   Compile with a checkpointer

   Checkpointing is local. The model swap does not change graph semantics, so existing threads keep working.

   checkpoint

   from langgraph.checkpoint.memory import MemorySaver
   
   graph = agent.with_config({"configurable": {"thread_id": "run-1"}})

3. 03

   Route cheap nodes to a fast tier

   Use a second client for classification and retrieval nodes so only planning hits the reasoning tier.

   two tiers

   fast = ChatOpenAI(model="deepseek-v4.1-flash", base_url="https://uncensoredai.dev/v1", api_key="...")

Commands and configuration below are shown exactly as the tool expects them, so they stay in English.