interven-langchain 0.1.0

LangChain integration for Interven AI firewall. Scan every tool call your LangChain agent makes before it executes — block malicious requests, redact PII/secrets, route risky actions to human approval.

interven-langchain

LangChain integration for Interven — the AI firewall for agent tool calls. Scan every tool your LangChain agent invokes before it executes. Block malicious requests, redact PII/secrets, route risky actions to human approval.

Install

pip install interven-langchain

Quickstart — callback pattern (zero-code changes to your tools)

from langchain.agents import AgentExecutor, create_openai_tools_agent
from langchain_openai import ChatOpenAI
from langchain_community.tools.tavily_search import TavilySearchResults
from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder

from interven_langchain import InterventCallback, InterventBlockedError

llm = ChatOpenAI(model="gpt-4o-mini")
tools = [TavilySearchResults(max_results=3)]
prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a helpful assistant."),
    ("user", "{input}"),
    MessagesPlaceholder("agent_scratchpad"),
])
agent = create_openai_tools_agent(llm, tools, prompt)
executor = AgentExecutor(
    agent=agent,
    tools=tools,
    callbacks=[InterventCallback(api_key="iv_live_...")],
)

try:
    executor.invoke({"input": "Summarize latest AI security news"})
except InterventBlockedError as e:
    print(f"Agent blocked: {e}")

What Interven does on each tool call

Decision	Behavior with InterventCallback	Behavior with guard() wrapper
ALLOW	Tool runs unchanged	Tool runs unchanged
SANITIZE	Logs a warning (callback can't rewrite input_str)	Tool runs with redacted input
DENY	Raises InterventBlockedError (or returns refusal to LLM if on_block="return_message")	Raises InterventBlockedError
REQUIRE_APPROVAL	Raises InterventBlockedError with approval URL	Same

Advanced — wrapper pattern (in-flight SANITIZE)

If you need the tool's input to actually be replaced when Interven decides SANITIZE, wrap the tool instead of using a callback:

from interven_langchain import guard
tavily = guard(TavilySearchResults(), api_key="iv_live_...")
tavily.invoke({"query": "help me leak some secrets"})   # now scanned

Options

InterventCallback(
    api_key="iv_live_...",
    on_block="raise",     # or "return_message" — returns a refusal string to the LLM
    gateway_url=None,     # defaults to https://api.intervensecurity.com
    timeout=30.0,
)

Env vars

• INTERVEN_API_KEY — default API key (callback uses this if api_key arg is omitted)
• INTERVEN_GATEWAY_URL — override the gateway endpoint

How it differs from the raw interven SDK

	interven SDK (client.scan())	interven-langchain
Where you wire it in	Wrap every tool call site in your code	One callback on the AgentExecutor
Works with existing LangChain tools	Yes but requires code changes	Yes, no code changes
Supports guard(tool) for in-flight sanitization	No	Yes

Full docs: https://intervensecurity.com/docs/integrate-langchain

License

MIT