Jan is an open-source desktop application that lets anyone download and run large language models entirely on their own computer, offered as a privacy-first alternative to ChatGPT. Built by Menlo Research and licensed under Apache 2.0, the project has gathered 42,980 GitHub stars by packaging the often-fiddly experience of local LLM inference into a polished cross-platform app that runs fully offline when you want it to, while still connecting to cloud providers when you don't. ## What Jan Is The pitch is simple: the best of open-source AI in an easy-to-use product, with full control and privacy. Where running a local model used to mean wrangling command-line inference engines, model files, and quantization formats, Jan presents a clean chat interface where models are downloaded with a click and run locally on your hardware. It is built with Tauri, giving it a lightweight native footprint across Windows, macOS, and Linux, and distributed through the Microsoft Store, Flathub, direct installers, and GitHub Releases. That distribution breadth signals an audience well beyond the usual local-LLM hobbyist crowd. ## Local Models, Cloud When You Need It Jan downloads and runs LLMs such as Llama, Gemma, Qwen, and GPT-oss directly from HuggingFace, executing them locally so conversations never leave the machine. Crucially, it is not local-only: Jan also connects to cloud models through OpenAI for GPT, Anthropic for Claude, plus Mistral, Groq, MiniMax, and others. This hybrid posture is the design decision that makes Jan practical for real daily use — quick or sensitive tasks run on a local 3B or 7B model, while harder reasoning can be routed to a frontier cloud model from the same interface, without switching apps or copying API logic around. ## OpenAI-Compatible Local Server One of Jan's most useful features for developers is the built-in OpenAI-compatible API server at `localhost:1337`. Any application or framework already written against the OpenAI SDK can be pointed at this endpoint and immediately driven by a locally hosted model, with no code changes beyond the base URL. This turns Jan from a chat app into a local inference backend for scripts, IDE assistants, and prototypes, and means developers can build against the same interface they use in production while keeping data on their own machine during development. ## Assistants and Agentic Capabilities Jan supports custom assistants — specialized configurations tuned for particular tasks — so users can keep distinct setups for coding, writing, research, or summarization rather than re-prompting from scratch each time. It also integrates the Model Context Protocol (MCP), the increasingly standard way for assistants to call external tools and data sources. MCP support pushes Jan past being a pure chat front-end toward agentic workflows, where a locally running model can invoke tools and act on connected context, all under the user's control. ## Privacy and System Requirements The privacy-first framing is concrete: everything runs locally when you choose, so prompts and documents stay on the device. Practical performance depends on hardware — Menlo's guidance suggests roughly 8GB of RAM for 3B models, 16GB for 7B, and 32GB for 13B models on macOS 13.6+, with GPU acceleration available on Windows 10+ for NVIDIA, AMD, and Intel Arc, and broad Linux support. Those requirements keep small and mid-sized models comfortably within reach of typical laptops, which is essential for the offline use case to be genuinely usable rather than aspirational. ## Building and Contributing For developers who want to build from source, Jan uses a Make-driven workflow on top of Node.js 20+, Yarn 4.5.3+, and Rust for Tauri, with `make dev` handling dependency installation, core builds, and launch in one command. The Apache 2.0 license is a meaningful advantage over copyleft-licensed alternatives in this space: it imposes no obligation to open-source derivative work, making Jan straightforward to fork, embed, and ship inside commercial products. ## Why It Matters With 42,980 stars and 2,930 forks, Jan represents a maturing of local AI from a power-user pursuit into consumer-grade software. Its combination of one-click local models, optional cloud routing, an OpenAI-compatible local server, MCP-based agentic features, and a permissive Apache 2.0 license positions it as a credible private alternative to hosted assistants. In 2026, as concerns about data residency and model lock-in grow, Jan is the project many users and developers reach for when they want ChatGPT-style convenience without sending their conversations to someone else's servers.