Local AI with no cloud calls¶

Understand's AI runs entirely on your own hardware by default — no code or prompts leave the machine unless you deliberately switch to a cloud provider. This page is the air-gapped recipe for keeping it that way. For the full model picker and the permissions detail, see Choose a model and What can the AI access? — this page just orchestrates them for a no-network deployment.

The rule: local model = no network, cloud provider = network¶

Understand's own Privacy page states the default plainly:

"By default, no code ever leaves your computer. … It uses a local model run entirely on-device, aside from a one-time model download, unless you switch to a third-party provider."

So there are exactly two things to control:

Keep the provider set to a local one — the bundled Understand Local model, or a local runtime (LM Studio / Ollama) on your own machine or network. These do inference on-device.
Handle the one-time model download on an air-gapped machine by placing the model file locally first (below).

If you switch the provider to OpenAI (ChatGPT), Anthropic (Claude), Google (Gemini), xAI (Grok), or a custom endpoint, prompts — including the code context in them — are sent to that provider. Don't do this in an environment that must keep code on-premises.

Handle the one-time model download¶

Bundled local models that aren't already installed are downloaded on demand from huggingface.co/SciTools/… the first time you select them. On a disconnected machine that download will fail, so do one of the following:

Use the model that ships with your install if it's already present, or
Pre-place the .gguf file: download it on a connected machine, transfer it to the target, and point Understand at it with Custom GGUF file in the model picker, or
Mirror the model internally and fetch it from there before going offline.

Once the model file is on the machine, inference is fully local and needs no network.

Keep local runtimes local¶

LM Studio and Ollama run on your machine (default endpoints 127.0.0.1:1234 and 127.0.0.1:11434) or on a host inside your own network. Choosing them keeps inference local. Just make sure the endpoint you configure is a machine you control, not a hosted service. Setup for both is covered in Set up a more robust local AI server.

Fully disable AI¶

If policy forbids running any model at all:

Turn off Enable AI on the Data preferences page (this also governs AI code search), and
For a hard guarantee, remove the bundled model file from the installation's conf/understand/models/ directory so there is nothing to load.

The MCP server — which lets external AI tools query a project — runs for every project you open. There is no per-project on/off toggle; the only per-project control is an opt-out checkbox, "Don't suggest this database for AI agent usage," which hides the database from agents but does not stop the server. It works entirely from your local .und database, so it makes no network calls of its own.

Confirm the rest of the machine is quiet¶

AI is only one network surface. To confirm nothing else reaches out (analytics, licensing, updates, crash reports), see What leaves the machine.