LangChain

Install and connect

pip install langchain-openai
export NOKEEP_API_KEY="sk-nokeep-test"  # your own key once it connects

import os
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="deepseek-v4.1-flash-uncensored",
    base_url="https://nokeep.ai/v1",
    api_key=os.environ["NOKEEP_API_KEY"],
)

reply = llm.invoke("Name three good names for a tavern.")
print(reply.content)
print(reply.usage_metadata)

In JavaScript the base URL goes inside configuration, not at the top level. A top-level baseURL is ignored and the call goes to the default host with your nokeep key.

Stream

import os
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="deepseek-v4.1-flash-uncensored",
    base_url="https://nokeep.ai/v1",
    api_key=os.environ["NOKEEP_API_KEY"],
    stream_usage=True,  # ask for the token counts in the last chunk
)

full = None
for chunk in llm.stream("Tell me a story in three lines."):
    print(chunk.content, end="", flush=True)
    full = chunk if full is None else full + chunk
print("\n", full.usage_metadata)

In Python, ChatOpenAI asks for usage on a stream only for its default host. Set stream_usage=True to get token counts with a custom base URL. The JavaScript client asks for them on its own. See Streaming.

Tool calling

from langchain_core.tools import tool

@tool
def get_weather(city: str) -> str:
    """Get the weather in a city."""
    return f"18°C and clear in {city}"

message = llm.bind_tools([get_weather]).invoke("What's the weather in Lisbon?")
for call in message.tool_calls:
    print(call["name"], call["args"])

Both models call tools. LangChain agents built on bind_tools work the same way. with_structured_output sends response_format with a JSON schema, which deepseek-v4.1-flash-uncensored takes and GLM drops. Use DeepSeek for structured output. See Tool calls and JSON output.

Thinking and samplers

import os
from langchain_openai import ChatOpenAI

thinker = ChatOpenAI(
    model="deepseek-v4.1-flash-uncensored",
    base_url="https://nokeep.ai/v1",
    api_key=os.environ["NOKEEP_API_KEY"],
    reasoning_effort="high",
    max_tokens=4000,
    extra_body={"min_p": 0.05},  # fields ChatOpenAI has no argument for
)

DeepSeek answers at once unless you ask it to think. Give a thinking call room: thinking counts toward max_tokens.

Gotchas

• `reasoning_effort` takes only the model’s own values. deepseek-v4.1-flash-uncensored takes none, low, high, max; glm-5.3-flash-uncensored takes low, high. Other apps’ minimal, medium and xhigh run at the model’s nearest level (medium as high, xhigh as max where the model has it); any other value gets 400 invalid_reasoning_effort. See Thinking.
• GLM picks its own tools. glm-5.3-flash-uncensored takes tool_choice auto or none only. Forcing a tool (required, or naming a function) gets 400 unsupported_tool_choice. Use DeepSeek when you need to force one.
• Fields a model can’t take are dropped, not refused. The call still answers, and the x-nokeep-ignored response header names what was left out. GLM keeps only tools, tool_choice and reasoning_effort, so its temperature and response_format are dropped. See which fields each model takes.
• Content parts are text and pictures only. image_url parts work on deepseek-v4.1-flash-uncensored; GLM reads text alone. Audio and file parts (a PDF) get 400 unsupported_input.
• `n` must be 1. Ask for several replies with several calls.
• Chat completions only. There is no /v1/responses and no /v1/embeddings; a framework that defaults to either needs to be pointed at chat completions, and embeddings need another provider.

ChatOpenAI(use_responses_api=True) and the JavaScript useResponsesApi: true call /v1/responses, which nokeep doesn’t have. Leave them off.