"Runs locally" has become a marketing phrase, and marketing phrases drift. One app means the model runs on your GPU but your prompts are logged. Another means the interface is local but the rendering is not. A third means everything is local until the update check, the analytics SDK and the crash reporter phone home.
If you are using local AI because your work is sensitive, "local" is a claim you should audit, not a feeling you should trust. Here is how, using our own app as the test subject.
Why This Matters More Than It Used To
The 2024 Cisco Data Privacy Benchmark Study, based on 2,600 privacy and security professionals across 12 geographies, found that 48% admit entering non-public company information into generative AI tools, and that 27% of organisations had banned generative AI at least temporarily over privacy and data security risks.
Visibility defaults matter too. Midjourney's own plan comparison states that Stealth Mode, "which keeps your images and videos private, is only available on the Pro and Mega Plans". That is a clear, documented policy. It is also a reminder that whether your work stays private can be a pricing tier someone else designed.
Check 1: Where Does the Engine Listen?
A local AI app almost always runs a local server that does the generating. The first question is who can reach it. A server bound to 127.0.0.1 is reachable only from your own machine. A server bound to 0.0.0.0 is reachable from your network.
Our answer: Vision Studio's engine is a bundled FastAPI service on 127.0.0.1:8000, reachable only from your own machine, and its generation API requires a token the app creates at launch. You can confirm the binding yourself with netstat -an on Windows or lsof -i :8000 on macOS and Linux.
Check 2: What Is Allowed to Leave?
Desktop apps built on Electron can restrict where their own interface may connect with a content security policy. If that policy allows any host, "local" is a hope.
Our answer: the interface is only allowed to connect to the engine on your own machine. Prompt tools, still images and video all start on the Local route.
Check 3: Is There Telemetry?
Analytics, telemetry and crash reporting are the most common quiet leaks, because they are usually third-party SDKs added for good reasons and forgotten. In a closed app you cannot check. In an open one you can search the source for them.
Our answer: the code contains no analytics, telemetry or crash reporting. The source is public at github.com/Git-Rocky-Stack/vision-studio, so that sentence is checkable rather than a promise.
Check 4: What Does It Go Online For?
No useful app is perfectly offline. The honest version of "local" lists every reason it touches the network. Be suspicious of any vendor that cannot produce that list.
Our list, with every route on Local:
| Network use | When | Your control |
|---|---|---|
| Model search and downloads | Only when you start them | Do not start them |
| Update check to updates.vision-studio-x.com | 15 seconds after launch, then every four hours | Set VISION_STUDIO_DISABLE_UPDATES=1 to switch it off |
| OpenRouter model list in Settings | Only if you saved an OpenRouter key | Do not save a key |
Downloads themselves are constrained: direct links are accepted only over HTTPS from CivitAI and GitHub, and are checked against a SHA-256 hash.
Check 5: What Happens When You Choose the Cloud?
Many local apps offer an optional cloud route for jobs too big for your GPU. That is fine, if the boundary is explicit. The question is exactly what crosses it.
Our answer: a cloud route exists only if you add your own OpenRouter or Hugging Face key and choose it, one job type at a time. Keys are encrypted with your operating system's secure storage. A cloud route sends prompt text and generation settings, never your images: a request that carries a mask, reference image, ControlNet guide or source image is refused on a cloud route. Cloud prompt tools can also include short excerpts of your earlier prompts as context, which the AI Director setting controls. A job too large for your GPU asks before it switches route.
What We Cannot Promise
An honest audit includes the limits, so here are ours.
- The update check is a network request. Like any request, it tells our delivery host that a copy of the app exists at your IP address. If that matters to you, switch it off with the environment variable above.
- Models come from third parties. Hugging Face and CivitAI see the downloads you start, under their own policies, not ours.
- Your operating system is outside our code. We can say what Vision Studio does. We cannot speak for OS-level telemetry, drivers or security software on your machine.
- A cloud route is a cloud route. Once you add a key and choose it, that provider's terms apply to what you send.
A vendor that lists no limits has not looked hard enough, or is hoping you will not.
The Five Questions, for Any Vendor
- What address does the local engine bind to, and does it require authentication?
- What hosts is the interface allowed to contact?
- Does the code include analytics, telemetry or crash reporting, and can I verify that?
- What is the complete list of reasons the app goes online, and can each be switched off?
- If I opt into a cloud route, exactly which data crosses the boundary?
A vendor that answers all five specifically, with source you can read, is making a claim you can test. A vendor that answers with "your privacy is important to us" is not.
Local Should Be Falsifiable
The point of local AI is control, and control you cannot verify is not control. Ask for the list. Read the source when it is available. Watch the network when it is not. "Runs locally" should be a statement you can try to prove false, and fail.
Vision Studio is free, open-source (MIT), and local by default. Every privacy and routing claim above is listed with the source files that implement it at vision-studio-x.com/features. Download it at vision-studio-x.com/download.
-Rocky
#Privacy #LocalAI #DataSecurity #Telemetry #OpenSource #AIGovernance #VisionStudio