Private AI assistants
These AI assistants do not use your prompts and files to train models by default, or run entirely on your own device. Compare chat retention and account requirements on each rating.
1. Jan Grade B
Open-source desktop app for running large language models offline on a local computer, with optional connections to remote AI APIs.
Score 89/100.
2. llama.cpp Grade B
Open-source C/C++ library and command-line tools for running large language model inference locally on CPUs and GPUs, including a built-in web server and chat interface.
Score 89/100.
3. Ollama Grade B
Open-source tool for downloading and running large language models on a local computer through a command line, API and desktop app. Optional cloud-hosted models are also offered.
Score 81/100.
4. Lumo Grade C
AI assistant from Proton that runs open-weight models on Proton's servers, with no logs of chats and zero-access encrypted chat history.
Based in Switzerland. Score 73/100.
5. GPT4All Grade C
Open-source desktop app from Nomic for running large language models locally, including chatting with local documents.
Score 72/100.
6. LM Studio Grade C
Desktop app for finding, downloading and running large language models locally, with a chat interface and a local API server.
Score 72/100.
7. Open WebUI Grade C
Self-hosted web interface for large language models that works with Ollama and OpenAI-compatible APIs, with multi-user support.
Score 72/100.
8. Brave Leo Grade C
AI assistant built into the Brave browser for chat, page summaries and writing help. Works without an account and can also use local or self-chosen models.
Based in the United States (Five Eyes). Score 61/100.
9. Duck.ai Grade C
AI chat from DuckDuckGo that relays prompts to third-party models without identifying the user, with no account needed and recent chats saved only on the device.
Based in the United States (Five Eyes). Score 60/100.
10. Kagi Assistant Grade D
AI assistant from Kagi that combines several third-party language models with Kagi Search results, available with a Kagi account.
Based in the United States (Five Eyes). Score 54/100.
11. Venice Grade D
AI chat and image generation service that proxies prompts to open and third-party models, keeping conversation history only in the browser.
Based in the United States (Five Eyes). Score 42/100.
How this list is made
Entries are ranked with our picks first, then by score against public criteria. Every answer links to evidence, and hosted services are tested automatically. Only entries answering no training on recordings with evidence are included. The list updates when ratings change.
Compare all: AI assistants. Criteria and scoring.
Questions
Does ChatGPT train on my chats?
See the ChatGPT rating for its current data controls and retention, with links to OpenAI's own documentation.
Can I run an AI assistant locally?
Yes. Local tools download open models and run them on your computer, so prompts never leave it.