Agents: the same rules, through a different door

An agent never opens your dashboard. Every control that lived in the interface is bypassed by construction. Not maliciously, just structurally.

Nexatron exposes a Model Context Protocol server over Streamable HTTP (JSON-RPC 2.0) secured with OAuth 2.1, so an MCP-capable client can ask governed questions of connected sources. Governance is enforced inside the same pipeline that serves the product's own interface, so an agent and a person asking the same question of the same source receive the same answer and both appear in the same audit log.

What it does

The agent inherits a human's reachPer-user authorization means what the agent can read is bounded by what the person it acts for could read. Not by what a shared integration account could read.
One pipeline, two doorsThe MCP path runs the same row-level security, the same certified metrics and the same semantic layer as the product interface. There is no second, looser path that exists because it was easier to build.
A tiered tool surfaceAnalytics tools are always on. The action-taking tools are a separate, entitlement-gated tier that is off by default, so read access and write capability are decisions you make separately.
One audit log for humans and agentsBoth land in the same hash-chained log. Two separate trails means nobody can answer "who saw this?" without joining them by hand.

What it deliberately won’t do

These are specifications, not caveats. Each one is a case where returning something would have been easy and wrong.

It will decline to an agent as readily as to a personAgents explore, so they ask for things that are not there more often than people do. The refusal path is the same one, and it names the boundary rather than returning something adjacent.
Cross-source questions need a certified linkAn agent cannot talk the platform into joining two systems on a key nobody approved. Where the link is absent, the answer is that it is absent.

Test the claim, don’t take it

Ask it something you already know the answer to, then ask it something your data cannot answer. Those two questions tell you more than a feature list.

How we measure

What it does

The agent inherits a human's reachPer-user authorization means what the agent can read is bounded by what the person it acts for could read. Not by what a shared integration account could read.

One pipeline, two doorsThe MCP path runs the same row-level security, the same certified metrics and the same semantic layer as the product interface. There is no second, looser path that exists because it was easier to build.

A tiered tool surfaceAnalytics tools are always on. The action-taking tools are a separate, entitlement-gated tier that is off by default, so read access and write capability are decisions you make separately.

One audit log for humans and agentsBoth land in the same hash-chained log. Two separate trails means nobody can answer "who saw this?" without joining them by hand.

What it deliberately won’t do

These are specifications, not caveats. Each one is a case where returning something would have been easy and wrong.

It will decline to an agent as readily as to a personAgents explore, so they ask for things that are not there more often than people do. The refusal path is the same one, and it names the boundary rather than returning something adjacent.

Cross-source questions need a certified linkAn agent cannot talk the platform into joining two systems on a key nobody approved. Where the link is absent, the answer is that it is absent.