What can the AI access? (permissions & privacy)

This page is for anyone who has to answer "is it safe to turn this on?" — especially in regulated or air-gapped environments.

By default, nothing leaves your machine

With the bundled local model (the default), Understand AI runs the model on your own hardware through an embedded llama.cpp runtime. Your source code, the prompts built from it, and the generated descriptions all stay local. Generated descriptions are stored in your project (ann_AI Overview.json) — nothing is sent to SciTools or any third party.

LM Studio and Ollama are also local: they run on your machine (or your own network host), so choosing them keeps inference local too.

What leaves the machine if you choose a cloud provider

If you switch the provider to ChatGPT, Claude, Gemini, Grok, or a custom endpoint, then each request — including the code context embedded in the prompt — is sent to that provider. Understand enforces some guardrails:

• An API key is required, and you must explicitly acknowledge data sharing and any fees before the provider setting is accepted.
• HTTPS is enforced whenever an API key is set — Understand refuses a non-HTTPS endpoint with a key.

If you have data-handling obligations, treat cloud providers as "code leaves the building" and choose a local model instead.

What the AI can touch inside Understand

The AI works from Understand's parsed model of your project and the source files in it. Descriptions and chat are read-and-explain operations — the AI does not modify your source code.

External AI agents connected through the MCP server have their own per-project permission model (read-only query tools; gated reparse, GUI, and file-read actions) — see Connect an AI agent via MCP.

Sharing descriptions without running the model

Because descriptions live in the project's ann_AI Overview.json, a machine that can't (or shouldn't) run a model can still read descriptions generated elsewhere by committing that file to version control. See overview & setup.

Fully disabling AI

Turn off Enable AI on the Data preferences page. For environments that must guarantee no model can run at all, remove the bundled model file from the installation's conf/understand/models/ directory so there is nothing to load.