Salus

Salus is a runtime authorization layer for AI agents that checks proposed tool calls against policy, session evidence, and intent before the backend executes them. It addresses the gap where an agent can hold valid credentials and permissions yet still take an incorrect action. Salus can hold unsupported writes, return the exact correction the agent needs, and record a decision receipt for each action.

The platform evaluates action identity, replay and duplicate control, input shape, content and data guardrails, time and rate limits, workflow state, evidence and facts, policy rules, confirmation and approval, and brokered execution. It supports Python SDK protection of tool callables, webhook routing for provider actions, and protection of custom tool executors, with compatibility across OpenAI, Anthropic, Gemini, LangChain, LangGraph, CrewAIAutoGen, Pydantic AI, LlamaIndex, Google ADK, Vapi, Retell, LiveKit Agents, Pipecat, MCP, HTTP, and computer use. Routes can run in shadow or enforcement mode. Decision records can be exported in CSV, JSON, agent JSON, or Markdown, with configurable retention.

Salus is for teams running agents that perform consequential actions such as writes, bookings, refunds, or external system calls. Pricing is based on registered agents and monthly action checks rather than seats, with unlimited seats on all plans. Three plans are offered: Pilot for first workflows, Partner for multiple agents and production workflows, and Enterprise for high-volume deployments, each with different agent and check allowances.