Demo

AI Connector Safety

AI-Powered

Give an AI assistant access to email, accounting and files, then decide what it may do. Watch reads run, sends wait for your approval, and a hidden instruction in a customer's email get caught. Or switch the guardrails off and see what leaves the building.

Pick a sample business

Its connectors, and what each may do

Read only blocks every write. Ask first shows you each send or payment before it happens. Allowed lets it run without asking.

Guardrails

Who's asking

A simulation: nothing leaves this page.

MCP, the Model Context Protocol, is how AI apps such as Claude and ChatGPT connect to tools. Once a connector can send or pay, the setup decides what the assistant may do.

Give each job only the access it needs. A summary only reads. Reminders need to send, with someone's approval.

Anything that sends, pays or deletes waits for a person, and the approval shows exactly what will go out.

Each connector acts as the person asking, with their access, never as a super-user.

Emails and documents can hide instructions meant for the assistant. A safe setup treats them as text to read, not requests to follow.

And it keeps a record of every action, including the ones it stopped.

This is a simplified simulation running in your browser. The businesses are made up, and every answer is pre-written.

A production setup adds: