Is a local AI agent private? What to check before connecting a model

Running an agent on your own computer can give you control over its installation, but “local” alone does not tell you where every piece of information goes. If that agent sends a prompt to a hosted model or calls an outside service, the receiving provider gets the data sent in that request. Start with the path your information takes, not the label on the app.

Map the journey of a task

Take a harmless example, such as asking for a summary of a project note. Identify where the source file lives, which agent reads it, which model processes the request, where the answer is stored and whether a connected service receives anything. Repeat for browser pages, attachments and messages: their paths may differ.

Before using real customer records or confidential documents, check your organization’s rules and each provider’s current terms, retention settings and account permissions. A local model may reduce the need to send prompts to a hosted model, but that does not automatically make every integration local or secure.

Look beyond model requests

An app may contact update services, send installation events or offer optional usage analytics. Connected integrations can have their own data practices. Find the product’s privacy and telemetry documentation, identify what is on by default, and test the available controls. Avoid promising “nothing leaves this computer” unless you have verified every relevant connection in your own setup.

What APX documents

APX’s privacy page says the app keeps chats, agents, files and keys on the computer. It also documents installation lifecycle events, optional aggregate usage reports and their controls. It says messages and agent output are not sent in those telemetry reports; that is separate from sending a request to a model provider or using an integration. The models guide describes hosted and local choices. Check your actual configuration and the destination provider’s terms.

The APX vs OpenClaw page describes both as options for running agents on your own hardware, with model choices and source links. It is a feature comparison, not a privacy audit. APX publishes this article and has an interest in the comparison.