Problem

Salus addresses the gap between an AI agent deciding to take an action and that action changing the real world. Agents can issue an incorrect refund without checking the order, send a sales email based on hallucinated data, or make an inaccurate backend update; traditional evaluations, output scoring, and observability generally identify these failures only after the fact. The economic pain is therefore tied to the blast radius of the action: lost money, damaged customer relationships, incorrect records, and operational or compliance exposure.

The clearest use case is protecting high-consequence tool calls such as refunds, payments, bookings, CRM writes, messages, claims, and human handoffs. Before an action executes, Salus can determine whether it is supported by the agent’s evidence and policy—for example, stopping a refund whose amount does not match the order or escalating an appointment change when the caller’s identity is ambiguous.

Product / Service

Salus is a B2B API and policy-aware proxy placed between an agent and the backend tools it calls. It is designed as a drop-in integration: teams change the tool endpoint rather than rebuilding or rewiring the agent, with support for OpenAI, Anthropic, LangChain, custom tool-calling agents, and voice stacks such as Retell and Vapi. Each proposed action is checked against prior tool outputs, conversation history, and policies expressed in YAML, Markdown, or plain English. Salus can allow, refuse, clarify, repair, rewrite, or escalate the action, returning structured feedback that the agent can use to retry.

The delivery model starts with shadow replay and shadow mode, allowing a team to see what Salus would have decided before enabling enforcement on a live action path. The product also includes synthetic red-team testing, decision logging, audit trails, and observability. Pricing is based on protected tool routes and action checks rather than seats or agent count, with a published pilot starting at $500 per month, a partner plan starting at $2,500 per month for up to ten routes, and custom enterprise deployment. Salus reports that 58% of blocked actions recover correctly in its benchmarks, while agents using Salus achieved up to 60% lower cost on one benchmark and 52% lower misalignment on another.

Market

Salus competes in the emerging AI-agent infrastructure and AI reliability market, specifically the runtime guardrails and pre-execution action-validation layer for agents that can affect customers, money, or operational systems. Relevant alternatives include Guardrails AI’s general AI reliability framework and native tool guardrails in agent platforms such as OpenAI’s SDK. LangSmith and Arize Phoenix are adjacent observability and evaluation platforms, while Lakera is an adjacent AI security provider. Salus’s proposed distinction is that it makes a decision on the action path immediately before execution and can repair or escalate the action, rather than primarily monitoring or evaluating behavior retrospectively.

The company appears to be at an early commercial stage. Y Combinator lists Salus as an active Winter 2026 company in Developer Tools, B2B API Infrastructure, and AI, founded by a two-person team; Salus’s own product page says it is “early” and invites teams to test the product against their agents in shadow mode. Public evidence shows a priced pilot, benchmark results, and anonymized design-partner traces, but does not disclose named customers, revenue, or production deployment volume. It is therefore best characterized as early and pre-scale, with pre-revenue status not publicly confirmed.