A runtime for agents that act. Palo Alto · May 13, 2026

Agents that act, safely.

A runtime that sits between every agent and every tool, deciding which actions ship and which get clarified, repaired, or held. One contract on the action path, across your entire stack.

Wires sent. Records written. Patients booked. Claims filed. Once an agent commits, you live with what it did. Salus is the runtime that makes commit-time the moment of judgment: every action checked against your policy and the evidence behind it, before it touches your customers, your money, or your record of the truth.

Book a demo →

Exhibit: Update Appointment escalated. Caller said Hi, this is John, move my appointment to Friday afternoon. Agent attempted update_appointment(customer_name=John, new_time=Friday 4:30 PM). Salus decision: escalate. Reason: multiple John records found; caller identity has not been verified. Backend would execute: No. — Exhibit A · identity.ambiguous_caller · appointment change, escalated for verification.

I. What Salus does

Three risks, three interventions.

Salus does not just block. It clarifies, repairs, rewrites, or escalates, then lets the corrected action through. Most actions ship. The dangerous ones ship better. Every decision is logged with the policy and evidence behind it.

Shadow analysis: 12 risky actions detected, 3 would have blocked, 9 would have escalated, 4 would have clarified, 12 would have prevented, avg latency 28ms.

Sample interventions, anonymized from design-partner traces.

II. How it ships

One endpoint change. Runtime safety, on.

Salus is a policy-aware proxy in front of every tool your agent already calls. Keep your stack. Change the URL. The agent does not know the difference. Your trust ledger does.

Diff of refund_client.py: change the request URL from api.acme.example/refunds to app.usesalus.ai/v1/proxy/refund-client and add an X-Salus-Proxy-Secret header.

You do not rebuild the agent. You change the endpoint.

III. From zero to enforcement

Four steps. One afternoon.

Connect your agent.

SDKs for OpenAI, Anthropic, LangChain, and custom tool-calling. Drop-in adapters for voice stacks like Retell and Vapi. No rewiring. No redeploy.

ii.

Import recent calls.

Salus extracts attempted actions, claims, handoffs. Builds a ledger.

iii.

Replay in shadow.

See what would have been allowed, clarified, rewritten, or escalated.

iv.

Promote one policy live.

Swap the tool URL. Salus is now on the hot path for that action.

IV. Works with

Your agent stack, untouched.

Reasoning

OpenAI, Anthropic, Gemini, any tool-calling LLM.

Orchestration

LangChain, LangGraph, CrewAI, LlamaIndex, your own loop.

Voice

Retell, Vapi, any audio loop.

Tools

MCP servers, REST, SDKs, internal services.

If your agent can call a tool, Salus can sit on that path.

V. Get in touch

If your agents touch anything real,
we should talk.

See Salus enforce a tool flow in five minutes, on whichever stack you already use. We will replay one of yesterday's sessions and show you the actions Salus would have decided differently.