The things you don't share
If you use ChatGPT, you've probably held something back: a made-up name instead of a real one, the numbers left out, a message you decided not to send. Most people do. In one study, 15 of 19 ChatGPT users cleaned up what they typed, and in a recent survey 44% of chatbot users said they'd share more if their chats weren't used for training.
Holding back has a cost: an answer can only be as good as what you were willing to tell it. The bank statement you didn't upload, the lab results you kept to yourself, the client file you knew you couldn't share, the question you never asked.
Etude is for those. Keep ChatGPT for everything else, and bring Etude what you'd never share there: your money, your health, personal matters, confidential work. The models Etude runs can read long documents, write, code and work through math, all on your own Mac, so none of it has to leave your computer.
Where your chats actually go
Typing into ChatGPT can feel as private as writing a note on your laptop. But your message crosses the internet to a data center, where the AI actually runs. The answer comes back; your message, and usually the whole conversation, stays there.
Nearly every mainstream assistant works this way: ChatGPT, Claude, Gemini, Grok, Meta AI, Copilot. Once your words reach a company's servers, what happens to them depends on its policies, its staff and, sometimes, a judge.
Kept, even after you delete them. In 2025 a US judge ordered OpenAI to preserve ChatGPT conversations, deleted ones included, as evidence in The New York Times's lawsuit. The order was lifted that October; the conversations saved under it were not.
Read by people. Contractors training Meta's AI read users' intimate conversations with the chatbot, and could see who those users were.
Used for training. Anthropic began training its AI on consumer chats, for everyone who didn't opt out.
Leaked. Conversations people thought were private keep turning up in public:
- BBC News · Aug 2025Hundreds of thousands of Grok chats appeared in Google search results.
- Fortune · Jul 2026A trove of conversations with Claude showed up in search results too.
- Reuters · Jan 2025A security firm found DeepSeek's sensitive data, including chat histories, exposed on the web.
- The Guardian · Sep 2026OpenAI said its agents leaked 53 images from ChatGPT users.
- BBC News · Feb 2026A Microsoft error exposed confidential emails to its Copilot AI.
- Ars Technica · Mar 2025Amazon began sending everything said to Echo devices to its servers.
None of these companies set out to expose the people using them. But once your conversation lives on someone else's computers, keeping it safe is their job, not yours, and sometimes it goes wrong. A conversation that was never sent to a data center can't leak from one.
What running AI “locally” means
ChatGPT isn't the AI itself: it's an app for reaching a model, which does the thinking. OpenAI keeps its leading models on its own servers, so ChatGPT has to send your words there.
Many capable models, though, are published for anyone to download: Qwen from Alibaba, Gemma from Google, even OpenAI's own gpt-oss. These open-weight models are often just a few gigabytes, and a recent Mac runs them comfortably. When the model runs on your own computer, your Mac's chip does the data center's work, and your words never have to leave. There is nothing to keep, read, train on, leak or hand over.
Cloud AI. The model runs in their data center, so your words go there.
Local AI. The model runs on your computer. Nothing needs to leave it.
Private by design, not by setting
Etude isn't the only Mac app that runs models locally; Ollama and LM Studio are well known too. But Ollama also offers cloud models in the same app: pick one marked “cloud,” and your conversation runs on Ollama's servers. Etude has no such switch. The model always runs on your Mac.
-
No account. There is nothing to sign up for and nothing to sign in to.
-
No cloud models. Nothing to pick by mistake, and no reason for us to nudge you toward paying for cloud tokens.
-
Works with Wi-Fi off. Once a model is downloaded, chats, documents and the agent all keep working offline. Wi-Fi Off? lists the few times Etude does go online, and what it sends.
-
Asks before it acts. Etude's agent reads the files in its workspace freely, but by default it asks before it edits one, runs a command, searches the web or opens a web page.
Fast, and free
Running AI locally used to mean slower answers. Etude runs on Pie, an open-source engine that grew out of research presented at SOSP 2025 by the team that builds Etude.
On Macs from the M2 to the M5 Max, Etude writes its answers as fast as or faster than MLX and llama.cpp, the engines most local AI apps rely on, and it often reads your prompt and documents much faster. The measurements are on the Compare page.
Already use an agent such as OpenCode, Hermes or the Pi coding agent? Etude can run it on its own engine, so you keep the tool you know and get Etude's speed and privacy. External Agents shows how.
Why is it free? Etude is made by Liszt AI, and it's how we show what Pie can do. So it will stay free of the usual catches: we promise that Etude will never require an account, never show ads, and never run your conversations on cloud models. If we ever offer a cloud product, it will be a separate app.
Your AI, on your computer
Keep ChatGPT for everything else. For your documents, your money, your health and your work, use an AI that never has to send them anywhere. Download Etude, pick a model, and turn off Wi-Fi if you like: your conversations stay where they belong, with you.
