They seem to have
You say Duck is trust-based? Isn’t NanoGPT even more so? They say they use TEE, but how do you actually verify that?
They’re both trust-based. It’s a similar model to a VPN, you’re just shifting trust. Though NanoGPT lets you pay with cryptocurrency so you at least have the ability not to connect your identity to a payment method like a credit card would
TEE only applies to certain models on NanoGPT. Not sure if there is a way to verify that on NanoGPT, but with MapleAI you supposedly can https://trymaple.ai/proof
I just had a pretty bad experience with my Maple account and deleted it. The model was claiming that it was ChatGPT 4. When I messaged support about it they said this is “common with open source models” which I haven’t experienced at all with open source models.
That’s quite common though. You can see the same in a lot of AI systems. AI is famous for hallucinating and this is a prime example. They could give it a system prompt where they define some facts but they probably don’t do that.
OpenRouter allows paying with cryptocurrency so it seems to have similar privacy guarantees as NanoGPT.
Just thought I’d mention here, maybe someone has but I missed it, Duck.ai actually does have one TEE model, gpt-oss. You can’t use the upload feature with it though, but you can search the web, which is nice.
I’m currently liking duck.ai for web searches and then Privatemode for things that I don’t need to search. I use Lumo sometimes with very specific prompts but I still just find it limited compared to these other options.
I’m trying to wean off Perplexity but its search is so good that I still use it with an anon email in incognito mode.
Don’t see a lot of talk about this generally.
I have tried Lumo and all the models available on duck.ai. The gpt-oss-120B and Haiku 4.5 with reasoning are the best IMO. I also like that gpt-oss is also likely not touching OpenAI servers at all, though I don’t know where the model is running.
I must say though I am very impressed with Ask Brave. I’ve not seen anyone recommend it. It uses a combination of Llama and Qwen IIRC. It integrates web-search really well and integrates it with Brave’s search indexing. IMO Brave also has the best search engine outside of Google. I have to say Ask Brave seems to be the best privacy-friendly cloud-based option available. It also doesn’t say no so much either like the better models on duck.ai.
What does the community think?
Claude AI with opt outs in a subscription, Claude AI with an API key, openrouter to GLM 5.2 with no data retention filter applied.
OpenAI is worse privacy policy wise and their opt outs are buried.
European AI models suck, i have 0 desire to use them personally right now.
I’ve tried everything you mentioned. Lumo is currently among the best out there. It’s multimodal, whereas duck.ai falls quite short in that regard. What’s more, I think GLM 5.2 (the model behind Lumo 2.0 Max) is one of the best models on the market right now.
I use Claude (Fable 5, Opus 4.6 and 4.8, and Sonnet 4.6 and 5) for work. These models remain the most reliable and give me the fewest headaches. Even so, for personal use, GLM 5.2 more than suffices. On top of that, NanoGPT offers GLM 5.2 TEE, if I recall correctly.
Lastly, I used to rely on Brave Leo and Ask Brave. They remain excellent alternatives, as you point out.
Private Cloud Compute. Anything else can’t prove that they really followed the privacy policy.
Two things are getting mixed together in this thread, and separating them.
One is whether the provider knows it is you. That is what a proxy like DDG actually fixes. The model host sees DDG’s infrastructure, not your IP or an account with your name on it. That part is real, and for most threat models it is the bigger half.
The other is whether what you typed gets stored. A proxy cannot fix that by itself. It comes down to the agreement between the proxy and the provider, and you have no way to check it from outside. Same for every no-logs claim here. They are promises, not guarantees, and that is fine as long as you know which one you are buying.
The only option that is different in kind is inference inside a trusted execution environment with remote attestation, where you can verify what is actually running instead of trusting a policy page. Model choice is limited and it is early, but it is the one place the answer is not just someone’s word.
So, copy-paste survives because it’s model agnostic. i keep the ide integration for code and a plain tab for everything else - getaskai.com, no login, fine for quick questions.
I use TEE models on nanoGPT because nanoGPT supports Monero. One of the perks is that I can also use Anthropic, open ai models whenever I’m willing to take the risk.
I tried Confer, but that wasn’t for me. There was no way to choose the model, and it was clear that Confer was trained to act like a psychotherapist. Great for those that need that, but that wasn’t what I was looking for.
I’ve settled on Tinfoil, after seeing it mentioned elsewhere here. It’s really good. Completely open source, exclusively uses open weight models. Everything uses TEE and, like Confer, Tinfoil uses a passkey for decryption.