Where your code actually goes when you build with AI
Most people do not ask this question until after they have pasted something they should not have. It is worth asking first, because the answer varies enormously between tools that look identical from the outside — a text box, a prompt, some code back.
The three places your code can end up
Every AI coding tool sits somewhere on a short spectrum, and the marketing rarely makes it obvious which.
- On a vendor's servers. Your project lives in their account. They hold it, they back it up, and their policy governs what may be done with it. This is the default shape for browser-based builders.
- On your machine, with a model call going out. The files stay local, but the text sent to the model — which includes your code — goes to a model provider over the network.
- Entirely on your machine. The model runs locally too, so nothing leaves the device at all.
These are genuinely different risk profiles, and "we take privacy seriously" on a landing page tells you which one you are in exactly never.
What local-first actually changes
Adrian is a desktop application rather than a website, so your projects and code sit in a folder you chose. Run it against a local model through Ollama and you are in the third case above: no account, no API key, and no network hop for the model call. Pull the network cable mid-build and the build carries on.
That is not a privacy setting you switch on. It is a consequence of where the software runs, which is the only kind of guarantee worth much.
The middle case, stated honestly
Bring your own API key and you are in the second case, not the third. Your files stay on your disk, but the prompt and the code context go to whichever provider issued that key, and their terms apply to it. Adrian takes no cut of that bill and nothing is proxied through us — but we would be misleading you if we called that setup private in the same sense as running locally. It is not.
For most work that trade is fine, and frontier models are meaningfully better at hard problems. It is worth knowing which trade you are making. What each option costs covers the money side of the same decision.
On training
We do not train AI models on your private projects, and we do not sell personal data. That is written into the privacy policy rather than living only in a blog post, which is where you should expect to find it for any tool — a claim that appears in marketing but not in the policy is not a commitment.
How to check any tool in five minutes
None of this is specific to Adrian. The same four checks work on anything:
- Find the privacy policy and search it for "train". If training on your content is permitted, it will usually be stated plainly, often with an opt-out you have to go and find.
- Ask where projects are stored. If the answer is an account on their platform, your code is on their servers regardless of what the homepage implies.
- Ask whether it works offline. A tool that cannot function without a network is sending something somewhere, by definition.
- Check who bills you for model usage. If the vendor bills you rather than the model provider, your prompts are passing through their infrastructure on the way.
The bottom line
There is no single right answer here. Hosted tools are convenient and often faster; local models are private and free but slower, and they struggle with long multi-file work. What matters is knowing which of the three cases you are in before you paste something that matters, not after.
Adrian is free on local models or your own key, so you can test the claim rather than take it. It is Windows only for now, and the current build is unsigned — Windows will show a SmartScreen warning on first run, and every release publishes a SHA-256 on the download page so you can verify what you got.