LangChain

Three ways to protect LangChain in Python, from least to most code. For LangChain.js, the callback handler is in the Node.js SDK guide.

Auto-instrumentation

LangChain’s chat models call the provider SDKs underneath, so auto-instrumentation scans them with no LangChain-specific code. ChatOpenAI goes through the openai package; ChatAnthropic through anthropic.

import promptguard
promptguard.init()  # reads PROMPTGUARD_API_KEY; call before creating models

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(model="gpt-5-nano")
llm.invoke("Hello!")  # scanned; raises PromptGuardBlockedError on a block in enforce mode

Inline guard in a chain

PromptGuardRunnable is a step you compose into an LCEL chain with |. Because it sits in the data flow, it can do what a callback cannot: on a redact decision it passes the redacted text on to the next step. On a block it raises PromptGuardBlockedError.

pip install "promptguard-sdk[langchain]"

from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI

from promptguard.integrations.langchain import PromptGuardRunnable

input_guard = PromptGuardRunnable().as_runnable()  # reads PROMPTGUARD_API_KEY
output_guard = PromptGuardRunnable(direction="output").as_runnable()

prompt = ChatPromptTemplate.from_template("Answer briefly: {question}")
chain = (
    {"question": input_guard}
    | prompt
    | ChatOpenAI(model="gpt-5-nano")
    | StrOutputParser()
    | output_guard
)
chain.invoke("What is prompt injection?")

A string, a dict with an input, query, question, text, content or prompt field, or a list of message dicts can be redacted in place. For any other value, a redact decision blocks in enforce mode rather than pass on content it could not rewrite. mode="monitor" logs instead of blocking or redacting.

Callback handler

PromptGuardCallbackHandler from promptguard.integrations.langchain does not run under current langchain-core, up to and including promptguard-sdk 2.3.0. It does not define the ignore_* attributes LangChain’s callback manager reads before every event, so the first callback raises AttributeError and the call fails. Use the inline guard or auto-instrumentation above until a release fixes it.

Proxy

Point the chat model at the proxy instead. Nothing runs in your process, and it works the same in any language.

import os
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="gpt-5-nano",
    base_url="https://api.promptguard.co/api/v1",
    api_key=os.environ["OPENAI_API_KEY"],  # forwarded to OpenAI
    default_headers={"X-API-Key": os.environ["PROMPTGUARD_API_KEY"]},
)

Get Started

Core Concepts

Gateway

Integrations

Shadow AI

Guides

Cookbooks

Platform

Going to Production

Resources

Roadmap / Preview

Auto-instrumentation

Inline guard in a chain

Callback handler

Proxy

​Auto-instrumentation

​Inline guard in a chain

​Callback handler

​Proxy

Auto-instrumentation

Inline guard in a chain

Callback handler

Proxy