Privacy & questions
Jarvis is built around local speech, local models and locally stored memory. These are the practical boundaries to understand before choosing your setup.
Jarvis is a private local AI assistant for macOS, Windows and Linux, created by Baris Sencan (isair). It combines voice and text conversation, a floating face, personal memory, dictation and tools. This website and the isair/jarvis repository describe this desktop assistant.
Voice recognition, chat, speech synthesis and saved memory can work offline once the models are downloaded and running locally. Initial installation and model downloads need a connection. Web search and weather use the internet, and connected tools have their own network requirements.
You choose where the models run. Keep every model server in your setup local if you want processing on your computer. If you configure a remote model endpoint, requests for that model go to that endpoint. An OpenAI-compatible server describes a supported protocol; it does not require OpenAI's cloud service.
Jarvis keeps a local diary and knowledge graph for saved conversations, preferences and plans. The Memory Viewer lets you browse them. Memory extraction and recall depend on the chosen models, so saved information can need correction.
While voice listening is enabled, a short, temporary transcript of nearby speech helps Jarvis understand the conversation when invited in. Pause Voice Listening from the menu to stop room transcription while keeping chat and supported dictation available.
Web features and connected tools can send information needed for a request. For example, weather needs a place and browser automation may load a website. MCP servers define their own access and behaviour. Choose servers and endpoints whose data handling fits your needs.
There is no single memory requirement for every setup. Chat models, speech recognition, quantisation and context length all contribute. The setup wizard shows a memory budget and guides local model selection. CPU fallback can be slower, and smaller models trade capability for lower resource use.
Yes. Jarvis supports Ollama and OpenAI-compatible servers. Setup lets you connect a server you already run, including compatible endpoints provided by LM Studio, oMLX or llama.cpp. Keep the server local if you want model processing on your computer.
Dictation is separate from the assistant conversation: hold a shortcut, speak and release to paste into another app. It is unavailable on macOS 26+. On supported macOS versions it needs Accessibility permission. Linux dictation requires X11, with limited Wayland support.
Jarvis's source is publicly available. Personal, educational, research and other non-commercial use is free under its custom non-commercial licence. Commercial use requires a separate licence from the copyright holder. The licence contains the full terms and contact details.
Browse the source, known limitations and published evaluation report. The report is dated, rather than a guarantee of accuracy for your model or hardware. You can also report an issue when something goes wrong.
Download the desktop app and choose where your models run.
Get started