Runtime control layer that evaluates every AI agent tool call before it executes

The design borrows from the API gateway pattern: agents keep their own reasoning, but nothing they decide to do reaches a database, a payment API, or a file system without passing a checkpoint. Polaxis runs a cascade rather than a single check. Regex scanning covers a set of injection and PII patterns, a risk scorer weighs a number of signals in under a millisecond, and an LLM semantic gate fires only on the small fraction of calls the cheaper layers cannot settle. Behind those sit a behavioral baseline meant to catch gradual drift, a session graph that looks for reconnaissance-to-exfiltration sequences, a per-agent threat level, and a policy engine holding the customer's own rules and spending caps.

Integration is a Python SDK wrapping tool calls, or an MCP proxy that requires no code change for clients such as Claude Desktop and Cursor; the SDK is framework-agnostic and documented against LangChain, LangGraph, CrewAI, AutoGen, PydanticAI, and the OpenAI Agents SDK. Blocked and escalated calls route to Slack or a dashboard for a human decision. Every evaluation is appended to a SHA-256 hash-chained log the company positions as tamper-evident and exportable for SOC 2 and EU AI Act auditors. Polaxis publishes performance and detection figures of its own, including sub-millisecond median latency on the pattern layers and a detection rate above 99 percent on its benchmark; those are the company's numbers. A free tier and an open-source SDK are available.