langchain-vaultak

Runtime security for LangChain agents, powered by Vaultak.

Intercept every agent action, score risk in real time, enforce policies, and automatically block dangerous behavior before it reaches your production systems — with two lines of code.

---

The Problem

LangChain agents can write to databases, send emails, execute code, and call external APIs. Without a security layer, a single misconfigured tool call can cause real damage — deleted records, leaked PII, unauthorized transactions.

langchain-vaultak wraps your existing LangChain agents with Vaultak's runtime security engine. Every tool call and agent action is scored and checked against your policy rules before it executes.

---

Install

pip install langchain-vaultak

Requires Python 3.8+, langchain-core>=0.1.0, and a Vaultak account at vaultak.com.

---

Quick Start

from langchain_vaultak import VaultakCallbackHandler

# Initialize with your Vaultak API key
handler = VaultakCallbackHandler(api_key="vtk_...")

# Pass as a callback to any LangChain agent
result = agent.run("Summarize our Q3 sales data", callbacks=[handler])

That is all. Every action the agent takes is now monitored, scored, and checked against your Vaultak policy rules.

---

What It Does

Event	Vaultak Action
Agent picks a tool	Risk-scores the action (0–10)
Tool call starts	Checks against policy rules
Tool call returns output	Scans output for PII and masks it
Tool or LLM error	Sends alert to your Vaultak dashboard
Chain error	Triggers automatic rollback
Risk score ≥ threshold	Blocks the action, raises exception

---

Configuration

handler = VaultakCallbackHandler(
    api_key="vtk_...",           # Required — your Vaultak API key
    agent_name="sales-agent",    # Label for this agent in the dashboard
    block_on_high_risk=True,     # Block actions that exceed the threshold
    risk_threshold=7.0,          # 0–10 scale — default 7.0
    verbose=True,                # Log all scored actions
)

Setting a lower threshold for sensitive workloads

# For agents with access to production databases or financial systems
handler = VaultakCallbackHandler(
    api_key="vtk_...",
    agent_name="prod-db-agent",
    risk_threshold=5.0,  # More conservative
)

---

Using with Different Agent Types

ReAct / AgentExecutor

from langchain.agents import AgentExecutor
from langchain_vaultak import VaultakCallbackHandler

handler = VaultakCallbackHandler(api_key="vtk_...")

agent_executor = AgentExecutor(agent=agent, tools=tools)
result = agent_executor.invoke(
    {"input": "your task"},
    config={"callbacks": [handler]}
)

LCEL (LangChain Expression Language)

from langchain_vaultak import VaultakCallbackHandler

handler = VaultakCallbackHandler(api_key="vtk_...")

chain = prompt | llm | output_parser
result = chain.invoke(
    {"input": "your task"},
    config={"callbacks": [handler]}
)

Global callback (applies to all chains)

from langchain.callbacks import set_handler
from langchain_vaultak import VaultakCallbackHandler

set_handler(VaultakCallbackHandler(api_key="vtk_..."))

---

Viewing Results

Every action scored by this handler appears in your Vaultak dashboard at app.vaultak.com. You can:

• See real-time risk scores for every tool call
• Review the full action history for any agent run
• Configure or tighten policy rules without code changes
• Set up alerts for high-risk events

---

Requirements

• Python 3.8+
• langchain-core >= 0.1.0
• vaultak >= 0.1.0
• A Vaultak account — sign up at vaultak.com

---

Links

• Vaultak docs: docs.vaultak.com
• LangChain docs: python.langchain.com
• Issues: github.com/vaultak/langchain-vaultak/issues

---

License

MIT — see LICENSE