Agent SDK

Behavioral security and policy enforcement for AI agents — one line of code.

The Quint Agent SDK captures every tool call your AI agent makes, enforces YAML-defined security policy at the tool-call boundary, and streams structured intent frames to the Quint cloud for real-time behavioral-security analysis.

Why

AI agents call tools — file reads, shell commands, API calls, database queries. Without visibility into what tools are called and with what arguments, you can’t detect prompt injection, data exfiltration, or unauthorized access.

The SDK sits at the tool-call boundary:

User Prompt → LLM → Tool Call → @quint.guard() → Tool Execution → Result
                                     ↓
                              Policy check (allow/block)
                              Telemetry capture (async)

What it captures

Event	Data
session_start	Agent name, framework, SDK version, metadata
tool_call	Tool name, arguments, timestamp, run ID
tool_result	Return value, duration, success/failure
policy_decision	Allow/block, matched rule, reason
session_end	Duration, final status, event count

Install

pip install quint-agent

With framework integrations:

pip install 'quint-agent[anthropic]'   # Anthropic Messages API
pip install 'quint-agent[langchain]'   # LangChain callback handler
pip install 'quint-agent[openai]'      # OpenAI function calling
pip install 'quint-agent[mcp]'         # Model Context Protocol
pip install 'quint-agent[all]'         # Everything

Quick start

import quint_agent

quint_agent.init(api_key="qak_...")

@quint_agent.guard()
def read_file(path: str) -> str:
    return open(path).read()

read_file(path="/etc/hosts")        # allowed — captured to cloud
read_file(path="/home/me/.env")     # blocked — ToolBlockedError

That’s it. Every call to read_file is now captured, policy-checked, and streamed to the Quint dashboard.

Quickstart

Get running in 2 minutes

Policy DSL

YAML rules for tool-call enforcement

Integrations

LangChain, Anthropic, OpenAI, MCP

API Reference

init(), guard(), Client, Policy