Skip to content
G

GUARDYX AI

Architecture

The control point between AI and the real world.

Your agents are on one side. The systems they can touch — money, customers, production, sensitive data — are on the other. Guardyx AI is the Control Plane in between. It checks every consequential action, decides whether to allow, hold, or deny it, and records the outcome — before any of it reaches those systems.

AI agents

  • Sales agents
  • Support agents
  • Finance agents
  • Copilots & research agents
  • Claude Desktop
  • ChatGPT Desktop

Making tool calls on your behalf.

Guardyx AI · Control Plane

  • Tool Gateway

    Every tool call arrives here first.

  • Policy Engine

    Decides allow, hold, or deny.

  • Approval Engine

    Routes exceptions to a human.

  • Audit

    Records every decision, tamper-evident.

Enterprise tools

  • CRM & email
  • Payment & ledger systems
  • Production infrastructure
  • Internal data & records

Nothing reaches these without passing the check.

AI agents on the left send every tool call to Guardyx AI in the middle, where the Tool Gateway, Policy Engine, Approval Engine, and Audit decide whether the action executes. Only permitted actions continue to the enterprise tools on the right.

Every AI action passes through Guardyx.

Three possible outcomes, every time

Allowed

In-policy actions execute immediately.

Held

Exceptions wait for a human sign-off.

Denied

Prohibited actions never reach the system.

Whichever way it goes, the decision is written down — so you can show, later and precisely, what your agents were allowed to do and what they were stopped from doing.

Where it runs

Hosted SaaS

We run it; you point your agents at it.

Self-hosted in your VPC / on-prem

The control point sits inside your own environment, next to the systems it protects.

Security & deployment detail

Want the technical version?

The developer docs cover the tool boundary, the MCP endpoint, the SDK, and how policies are expressed and evaluated.

Put a control point in front of every consequential action.

One integration point in front of your tools, so the same rules apply no matter which agent or framework made the call.