janhq/jan

Jan is a desktop chat client positioned as an open-source ChatGPT replacement. Its distinguishing feature is that it can run models entirely on your own machine, while also allowing cloud models when you prefer them. It is aimed at people who want a familiar chat interface over local or hosted models, not at developers building applications on top of a model.

What it does

You can download open models such as Llama, Gemma, Qwen and GPT-oss from Hugging Face and run them locally through the bundled llama.cpp engine. Alternatively, you can connect the same client to hosted models from providers including OpenAI, Anthropic, Mistral, Groq and MiniMax. The privacy claim is conditional, not universal: the README says everything runs locally "when you want it to," so data stays on-device only for the local-model path. Cloud-backed conversations follow that provider's handling instead.

Beyond plain chat, Jan supports custom assistants tailored to particular tasks, and it exposes an OpenAI-compatible server on localhost:1337 for other applications. It also offers Model Context Protocol integration for agentic capabilities. These are secondary features for people who want to wire Jan into a broader workflow; the core experience remains a chat client.

Requirements

Using the packaged app is the intended path: installers are listed for Windows, macOS (universal), and Linux as deb and AppImage, with Microsoft Store and Flathub options also provided. An Arm64 Linux build is not offered as a standard download; the README points to a separate how-to instead. Minimum guidance is macOS 13.6+ and Windows 10+, and RAM needs scale with model size — about 8GB for 3B models, 16GB for 7B and 32GB for 13B. Windows also lists GPU support for NVIDIA, AMD and Intel Arc.

Building from source is a different and much heavier proposition, requiring Node.js 20+, Yarn 4.5.3+, Make, Rust, and on Apple Silicon an extra Metal toolchain download. Windows source builds need Git Bash plus Visual Studio 2022 Build Tools, LLVM, Ninja and CMake, with CUDA Toolkit only for CUDA engine variants. The README documents a specific Windows path-length failure during llama.cpp compilation and the workaround of a short build directory. None of this is needed to use the installed app.

Practical notes

The README states Apache 2.0, though the repository's licence field is marked unrecognized, so confirm the actual licence file before redistributing. There are no stated benchmarks in the supplied material, and no documented data-migration or upgrade behavior; installation guidance and troubleshooting are the places to check for compatibility details and known failure modes.

HotRepo's take

The project was observed once on the daily trending board on 2026-10-30 at 45,163 stars, up 34 that day. One recorded day is not enough to describe momentum in either direction.

Project source and documentation