“GDPR-compliant” is printed on almost every AI tool. What that means: there's a contract. Local means something else: your data doesn't leave. The difference isn't a legal one, it's a technical one.
If you want to make progress on data protection, you need to tell these three setups apart. Marketing likes to blur them together.
| Setup | Where your data goes | What data protection law says |
|---|---|---|
| Cloud tool in the browser | To the provider, where it's processed and usually stored | Requires a data processing agreement (DPA), often a transfer of data to third countries outside the EU/EEA, and you have to take the provider's word for it |
| App on your computer, model in the cloud | The task and everything the agent reads for it go to the model provider. Whatever it doesn't need stays on your computer | A much smaller attack surface: there's no account where everything comes together. The model provider still needs a legal basis |
| App and model on your computer | Nowhere. The chain ends on your device | No transfer, so no third-country question either. The simplest case there is |
nebra AI works in setup two or three, and you decide during setup. Setup two is the standard case: the app runs on your Mac and sends each individual task to the model you already have an account with. Setup three is the strict case: a model that runs on your computer, with no internet.
Even in setup two, everything the agent reads for a task leaves your computer: the instructions, the pages involved and the knowledge it uses for them. If you don't want that, you need setup three, and that costs you speed and writing quality. A local model on a laptop doesn't play in the same league as the big providers.
And one more thing: no tool makes you GDPR-compliant. That depends on which data you process, why, and what's in your records of processing activities. Software can make the work easier, it can't do it for you.
In nebra AI, this route is called Qwen via Ollama. It's an open language model that runs as a service on your Mac.
Who it's worth it for: anyone who works with health data, HR data, client files or engineering documents. There, the question isn't whether a contract is good, it's whether the data leaves the building. The answer has to be no, and this setup can keep it that way.
You don't have to decide up front. The setup assistant shows you all four AI options, and you can switch later.
That question can't be answered as asked, and anyone who answers it with a flat yes is selling you something. GDPR compliance depends on which data you process and on what legal basis. What we can say: the app runs on your computer, we don't store your company knowledge, and with the local model no text leaves your device.
For working with the local model, yes: Qwen runs via the Ollama service on your Mac. To connect your website and publish changes, you obviously need a connection to that website. License checks and updates need a connection to our server now and then.
It makes sense from 8 GB of RAM upward, and the app picks the model size based on your RAM. With more memory, a larger model runs and the results get better.
Speed and writing quality. A model running on a laptop is weaker than Claude or GPT. It also works in text mode: no follow-up questions while it works, no approval cards, and a conversation can't be continued.
Yes, if you use a cloud model and the agent reads customer data for the task. Then everything it needs for that task goes to the provider. Your credentials, experts and company knowledge are stored on your computer, not with us. If even the task texts must not leave your company, use the local model.
The setup assistant shows you all four AI options, from the recommended Claude account to the local model with no sign-up.
Written by
Founder and managing director of konversion.digital, a behavioral economist with 20 years of hands-on conversion work for mid-sized businesses and more than 800 A/B tests run. Author of the German book “Zeiträuber & Eroberer” and a keynote speaker on conversion, AI and the psychology of decision-making. Nebra AI packages his methodology so teams can apply it without an agency.