Top AI Product

An AI agent that can process refunds or change customer records is useful right up until it does the wrong one. Salus, a YC W2026 startup, is built for exactly that risk: a runtime guardrail layer that checks every action an agent takes before it executes.

## How Salus works

Salus is a policy-aware proxy. You change the URL your agent hits — instead of calling OpenAI or Anthropic directly, traffic routes through Salus — and it intercepts every tool call before execution. Each call is validated against versioned policies and an evidence cache; if something violates policy, Salus blocks it and returns structured feedback so the agent can retry correctly rather than fail blindly. Because it sits at the proxy layer, it works with any tool-calling agent, including OpenAI, Anthropic, LangChain, custom stacks, and voice platforms like Retell and Vapi.

## Who needs it

The target is companies running autonomous agents in production where mistakes carry financial consequences — refund processing, customer operations, anything touching real systems. Founded by five ex-Stanford engineers and backed by $4M, Salus is positioning itself as the safety layer underneath any agent that acts in the real world, separating “the agent decided” from “the action was allowed.”