langchain-concentrate

LangChain integration for Concentrate.ai — a unified LLM gateway that lets you access OpenAI, Anthropic, Google, DeepSeek and other providers through a single API, with governance, spend tracking, fallbacks and auto-routing.

Concentrate exposes an OpenAI Responses API–compatible endpoint at https://api.concentrate.ai/v1, so this package is a thin subclass of langchain-openai's ChatOpenAI that pins the base URL, forces the Responses path, and reads its key from CONCENTRATE_API_KEY.

Install

pip install langchain-concentrate
# or
uv add langchain-concentrate

Quickstart

export CONCENTRATE_API_KEY="sk-cn-v1-..."

from langchain_concentrate import ChatConcentrate

model = ChatConcentrate(model="gpt-5.4")
print(model.invoke("Say hi in three words.").text)

Streaming

for chunk in model.stream("Count from 1 to 5."):
    print(chunk.text, end="", flush=True)

Auto-routing

Pass model="auto" with a routing directive to let Concentrate pick a provider based on your policy:

cheap = ChatConcentrate(
    model="auto",
    routing={"strategy": "min", "metric": "cost"},
)
cheap.invoke("What is 2 + 2?")

Tool calling & structured output

Because ChatConcentrate inherits from ChatOpenAI, bind_tools and with_structured_output work out of the box for models that support them downstream.

from pydantic import BaseModel

class Answer(BaseModel):
    number: int

structured = ChatConcentrate(model="gpt-5.4").with_structured_output(Answer)
result = structured.invoke("What is 2 + 2?")
assert result.number == 4

Configuration

Any keyword accepted by ChatOpenAI — temperature, max_output_tokens, timeout, default_headers, extra_body, model_kwargs, etc. — can be passed through.

Development

uv sync --all-groups
make test              # unit tests (no network)
make integration_tests # against real Concentrate, needs CONCENTRATE_API_KEY

License

MIT

Disclaimer

Portions of this package were generated with the help of an AI coding assistant. All code is human-reviewed.