LangChain¶

LangChain's chat models call the provider through the official SDKs, so they take a base URL. Point it at Privyx and every prompt, chain, and agent step is masked on the way out and restored on the way back.

OpenAI models¶

privyx proxy --upstream https://api.openai.com

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(model="gpt-4o-mini", base_url="http://localhost:8000/v1")

print(llm.invoke("Write a short greeting to alice@example.com").content)

The model receives <PRIVYX_EMAIL_1>; content has the address back. Streaming works the same way:

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(model="gpt-4o-mini", base_url="http://localhost:8000/v1")

for chunk in llm.stream("Write a short greeting to alice@example.com"):
    print(chunk.content, end="", flush=True)
print()

Anthropic models¶

privyx proxy --upstream https://api.anthropic.com

from langchain_anthropic import ChatAnthropic

llm = ChatAnthropic(model="claude-opus-5-5", base_url="http://localhost:8000")

print(llm.invoke("Write a short greeting to alice@example.com").content)

The Anthropic base URL has no /v1.

Without changing code¶

Both classes read the SDK's environment variable when no base_url is given:

export OPENAI_BASE_URL=http://localhost:8000/v1
export ANTHROPIC_BASE_URL=http://localhost:8000

Good to know¶

Tools and agents. Tool-call arguments are restored before LangChain runs the tool, so your tools receive real values, and their results are masked again on the next request.
Embeddings are not masked. OpenAIEmbeddings calls /v1/embeddings, which Privyx forwards as sent. If a vector store must not hold original values at the provider, mask the text first with privyx mask, or refuse the path with proxy.passthrough_unknown: false.
One mapping per conversation. Pass a session id with default_headers={"x-privyx-session": "..."}; see One mapping per user or conversation.