What it does Jan is a desktop chat client built with Tauri that runs LLMs locally via llama.cpp while also letting you plug in OpenAI, Anthropic, Groq, and other cloud providers. It downloads models from HuggingFace, spins up an OpenAI-compatible local server at localhost:1337, and recently added MCP support for agentic workflows. Basically: one UI for both “my GPU is melting” and “I need GPT-4 for this one.”

The interesting bit The project treats privacy as a toggle, not a religion. Everything can run offline, but you’re not locked out of cloud models when you need them. That’s a harder design problem than it sounds — most local-first tools make cloud feel like a betrayal, and most cloud tools make local feel like a hack.

Key highlights

• Ships as native desktop apps for Windows, macOS, and Linux (including AppImage and Flathub)
• Local model support through llama.cpp with GPU acceleration on NVIDIA, AMD, Intel Arc, and Apple Silicon
• OpenAI-compatible API server so other tools can piggyback on your local instance
• Custom assistants and MCP integration for task-specific workflows
• 42K+ stars and active development with multilingual docs

Caveats

• RAM requirements escalate fast: 8GB for tiny 3B models, 32GB for 13B
• Linux ARM64 support exists but requires a manual workaround per a GitHub issue
• The “privacy first” claim holds for local mode; cloud integration is opt-in but still sends data to third parties

Verdict Worth a look if you want a polished desktop alternative to juggling Ollama in a terminal and ChatGPT in a browser. Skip it if you’re already happy with a terminal + API key workflow, or if your machine has less than 16GB RAM.