AgentSentinel

Safety controls for AI agents — budgets, approvals, rate limits, audit logs, and DLP.

Installation

pip install agentsentinel

Quick Start

from agentsentinel import AgentPolicy, AgentGuard

policy = AgentPolicy(
    daily_budget=10.00,
    require_approval=["send_email", "delete_*"],
    rate_limits={"web_search": "10/min"},
)

guard = AgentGuard(policy)

@guard.protect("web_search")
def search_web(query: str) -> str:
    return f"Results for: {query}"

Features

• 💰 Budget Controls — Daily/hourly spending limits with per-model caps
• ✋ Approval Gates — Require human approval for sensitive operations
• ⏱️ Rate Limiting — Prevent runaway loops with sliding window limits
• 📋 Audit Logging — Complete trail of every tool invocation
• 🔒 DLP & PII Detection — Block credit cards, SSNs, API keys from leaking
• 🌐 Network Controls — Allowlist/blocklist outbound domains

Framework Integrations

# LangChain
from agentsentinel.integrations.langchain import protect_langchain_agent
executor = protect_langchain_agent(executor, policy=policy)

# CrewAI
from agentsentinel.integrations.crewai import protect_crew
crew = protect_crew(crew, policy=policy)

# LlamaIndex
from agentsentinel.integrations.llamaindex import protect_agent
agent = protect_agent(agent, policy=policy)

# OpenAI Assistants
from agentsentinel.integrations.openai_assistants import protect_function_map
functions = protect_function_map(functions, guard=guard)

# Anthropic Tools
from agentsentinel.integrations.anthropic_tools import protect_tool_handlers
handlers = protect_tool_handlers(handlers, guard=guard)

101 Models Supported

Cost tracking for OpenAI, Anthropic, Google, Mistral, Cohere, Meta, AWS Bedrock, Azure OpenAI, Groq, Together, Perplexity, DeepSeek, xAI, and local models.

Documentation

Full docs at docs.agentsentinel.net

License

MIT