Validate all your agent's actions before they execute
Block incorrect actions at runtime and provide immediate feedback to guide retries.
Our Product
01
Runtime Guardrails
Intercept and validate every tool call at execution time. Block actions that violate policy or aren't grounded in evidence before they execute.
02
Self-Repair
When an action is blocked, Salus returns structured feedback explaining exactly what failed, enabling the agent to self-correct and retry.
03
Full Visibility
Watch every agent interaction stream in real-time. Full traces, token usage, and latency breakdowns.
04
Evals
Generate thousands of adversarial and realistic scenarios to measure tool-call correctness before deployment.
from salus import Session, PreflightBlockedError
from openai import OpenAI
client = OpenAI()
session = Session(enforce=True)
@session.protect(produces_evidence=True)
def lookup_order(order_id: str):
return db.get_order(order_id)
@session.protect(is_commit=True, requires=["lookup_order"])
def issue_refund(order_id: str, amount: float):
return payments.refund(order_id, amount)Easily Integrate
One decorator per tool call. That's it.
Works with OpenAI, Anthropic, LangChain, LangGraph, and CrewAI.
Ready to secure your AI agents?
Book a 30 minute live demo with our team to learn how Salus can protect your business from costly mistakes.
Got questions? Reach out to us anytime: founders@usesalus.ai
Security
Report a Security Issue
Customers can report vulnerabilities or security concerns directly to our security team.
Email: security@usesalus.ai
Support: support@usesalus.ai
Reports are triaged and tracked in our internal ticketing system.