How it all fits together.

Start here

NrtVix AI has two parts that work together. The Agent Builder lets you create an AI agent that talks to customers and asks to take actions. The Guardian decides whether each action is allowed, and a trusted gateway carries out the ones that are.

The most important idea is the separation: the agent proposes, the Guardian decides, the gateway executes. No single part can do all three, so a mistake or a manipulated prompt in one part cannot become an action on its own.

Workspaces

A workspace is your business inside NrtVix AI. It holds your agents, your policy, your budgets, your approvals and your audit trail. Workspaces are isolated: an agent in one workspace cannot read or act on another, and every lookup is scoped to the workspace it belongs to.

Agents

An agent is one configured assistant. It has a name, business information, a list of tools it may use, and a monthly budget. A hosted agent is created in the browser. An external agent is one you built elsewhere and connected through the SDK.

An agent has an identity but no usable secret. It cannot hold a payment key, and it cannot grant itself a permission.

Tools

A tool is one kind of action, with a strict list of arguments. Today there are four: look up a customer, refund an order, add a ticket note, and prepare a customer export. An agent can only use the tools you ticked when you created it; the others do not exist for it.

Decisions

Every tool call gets exactly one decision. Allow means it runs at once. Review means it waits for the owner. Block means it does not happen. The decision depends on your policy, the amount or size of the request, and the agent’s remaining budget, never on how the request was worded.

Approvals

When a request needs review, it appears in the owner’s approval queue with the exact action, target and amount. Approving it creates a single-use permission bound to that exact request. If anything about the request changes afterwards, the permission does not fit and the action does not run.

Receipts and audit

Each executed action leaves a receipt. Each decision, approval and stop is written to an audit trail in which every entry includes a hash of the one before it. Changing or removing an old entry breaks the chain, and a verification check reports where.

Where to go next

• Guides: writing business information, choosing limits, and understanding prompt injection.
• API reference: endpoints, streaming events and errors.
• SDK: the guarded toolbox for Python.
• Security: the model behind all of this.

Preview. Everything on this page describes the current build. Names and paths may change before release, and the hosted API is not live yet. Documentation opens fully with early access.