Runtime API that inspects and blocks AI agent actions before they execute, then gives the agent structured feedback for safe retries.
Salus is a runtime safety layer that wraps AI agent inference calls to check proposed actions against explicit policies and evidence before the backend executes. It intercepts unsafe writes, issues decision receipts for every intervention, and provides structured feedback so the agent can correct and retry. A shadow-first rollout approach lets teams replay traces, author policies, and run alongside production before promoting to live enforcement. Compatible with OpenAI, Anthropic, Gemini, LangChain, LangGraph, CrewAI, LlamaIndex, Retell, Vapi, and MCP servers. Salus is a product of Salus.