SDKs and LangChain
Server code should use the official OpenAI SDK, or any library that accepts a custom Base URL. Put the key in an environment variable or a secret store. Do not commit it.
export OPENBRIDGER_API_KEY="sk-..."
export OPENBRIDGER_BASE_URL="https://openbridger.com/v1"
export OPENBRIDGER_MODEL="MODEL_ID_FROM_CONSOLE"
Node and Python OpenAI SDK samples are on OpenAI-compatible calls.
Before production
- Use a different key for development, test, and production.
- Set connect and read timeouts. Retry
429and temporary upstream 5xx a limited number of times with backoff. Do not retry401or a403for insufficient credit. - Log the request id, model, status, and latency. Drop the key and sensitive prompt text.
- Put your own limit on input length and
max_tokens. The gateway rejects some illegal values. That is not your quota policy.
LangChain (Python)
import os
from langchain_openai import ChatOpenAI
model = ChatOpenAI(
model=os.environ["OPENBRIDGER_MODEL"],
api_key=os.environ["OPENBRIDGER_API_KEY"],
base_url=os.environ["OPENBRIDGER_BASE_URL"],
)
print(model.invoke("Hello OpenBridger").content)
The package is langchain-openai. It uses the OpenAI SDK, so base_url still includes /v1.
LangChain (JavaScript)
import { ChatOpenAI } from "@langchain/openai";
const model = new ChatOpenAI({
apiKey: process.env.OPENBRIDGER_API_KEY,
configuration: {
baseURL: process.env.OPENBRIDGER_BASE_URL,
},
model: process.env.OPENBRIDGER_MODEL,
});
const response = await model.invoke("Hello OpenBridger");
console.log(response.content);
This baseURL is also https://openbridger.com/v1. If a package version moved the field to a top-level baseURL, follow that version. The finished request host should be openbridger.com and the path should end in /v1/chat/completions with a single /v1.
Common problems
| What you see | What to do |
|---|---|
| Requests still go to official OpenAI | The environment variable is not in the process, or the parameter name is wrong |
| Browser CORS | Move the call to your server |
| Tool calls fail | invoke plain text first, then check whether the model supports tools |
| LangChain wraps a credit error as a generic exception | Print status and the body, then use Error codes |