LangChain Custom API Setup

LangChain's official Python integration for OpenAI lives in the separate langchain-openai package. Connect TokenHub by passing a custom base_url to ChatOpenAI. This is a code integration; there is no App settings screen.

Step 1: install the official integration package

pip install -U langchain-openai

For a project managed by uv, use uv add langchain-openai. Installing the base langchain package alone does not provide ChatOpenAI.

Step 2: instantiate ChatOpenAI explicitly for TokenHub

export TOKENHUB_API_KEY="sk-..."

import os
from langchain_openai import ChatOpenAI

model = ChatOpenAI(
    base_url="__API_BASE_URL__/v1",
    api_key=os.environ["TOKENHUB_API_KEY"],
    model="YOUR_TOKENHUB_MODEL_ID",
    timeout=60,
    max_retries=2,
)

response = model.invoke("Reply with one sentence to confirm the connection.")
print(response.content)

Base URL resolution favors the constructor's base_url / openai_api_base, then OPENAI_API_BASE, followed by the underlying SDK's OPENAI_BASE_URL. Passing it explicitly prevents an old OpenAI URL in the project environment from overriding TokenHub.

Step 3: verify one capability at a time

1. Run the minimum invoke call to confirm normal Chat Completions first.
2. Add stream only when basic invocation works. When OPENAI_BASE_URL points to a third-party endpoint, LangChain does not automatically enable stream_usage; set stream_usage=True only if the TokenHub model and endpoint support streaming usage.
3. For an agent, bind one minimal tool with bind_tools and inspect tool_calls on the returned object. Do not treat text that merely describes a call as a tool call.
4. Test structured output, images, and reasoning only after those layers work.

ChatOpenAI explicitly targets fields in the OpenAI API specification. Provider-specific extensions such as reasoning_content, reasoning, or reasoning_details may not be extracted or preserved. A successful text response does not prove that those extensions work.

Troubleshoot at the LangChain layer that failed

• Import fails: confirm langchain-openai is installed.
• 404: inspect the effective base_url, including /v1, and check for environment-variable overrides.
• Model error: use the complete TokenHub model ID in model.
• Streaming text works but usage is absent: inspect stream_usage and whether the endpoint supports stream_options.include_usage.
• Agent tool fails: isolate it with bind_tools; a successful basic invoke is not a tool test.
• Non-standard reasoning field is absent: use the provider-specific LangChain integration or process the raw response directly.

Official reference

• LangChain: ChatOpenAI integration