Jan
An open-source ChatGPT alternative that runs fully offline — a polished desktop app over llama.cpp with a model hub, MCP support, and a local API server.
Jan (janhq/jan, Apache-2.0, ~43k stars, by Menlo Research) is the open-source answer to LM Studio: a Tauri desktop app that downloads and runs local models via a llama.cpp engine, exposes an OpenAI-compatible API on localhost:1337, supports MCP for agentic use, and optionally connects cloud providers with your own keys. 100% offline-capable; 5.7M+ downloads.
Jan is the open-source desktop app for local AI — "an open-source ChatGPT alternative that runs 100% offline," in the project's own words. Built by Menlo Research as a Tauri (Rust) app over a llama.cpp engine, it wraps model discovery, chat, a local API, and MCP into something a non-terminal user can love — while staying Apache-2.0 all the way down.
Highlights
- Model hub built in — browse and download open-weight models (Llama, Gemma, Qwen, gpt-oss, …) from Hugging Face inside the app.
- OpenAI-compatible local API on
localhost:1337— other tools and agents target Jan like any provider. - MCP support — stable since v0.6.9 (August 2025), making Jan a local host for Model Context Protocol tooling.
- Cloud as an option, not a default — connect OpenAI/Anthropic/Mistral/Groq with your own keys alongside local models.
- Active engine work — llama.cpp auto-tuning (v0.7), a unified router with multi-token prediction (v0.8), AMD ROCm on Linux (v0.8.2, June 2026).
- Genuinely open — Apache-2.0 (relicensed from AGPL in May 2025), ~43k stars, 5.7M+ downloads.
In an AI-assisted workflow
Download from jan.ai, pull a model that fits your hardware (the quantization literacy applies), and chat — or flip on the local server and point your BYO-model tools at localhost:1337. The privacy story is the cleanest in the desktop class: with local models, nothing leaves the machine.
TIP
Jan + MCP is an underrated combo for a fully-local agent playground: local model, local tools, zero cloud surface — useful both for sensitive work and for understanding agent mechanics without an API bill.
Good to know
Runs on macOS 13.6+, Windows 10+, and Linux (deb/AppImage, Flathub, Microsoft Store). One citation quirk: GitHub's license API reports "Other" because of a custom copyright header — the LICENSE text is standard Apache-2.0. There's no hosted/cloud Jan; it's desktop-first by design. Where it sits against LM Studio's polish and Ollama's headless ubiquity is mapped in Best Tools for Running LLMs Locally.
Frequently asked questions
- Jan vs LM Studio — what's the real difference?
- Licensing and philosophy. Both are polished desktop apps for running local models on a llama.cpp foundation. Jan is fully open source (Apache-2.0, relicensed from AGPL in May 2025), community-driven, and privacy-framed ('answers only to you'); LM Studio is proprietary freemium with arguably more hardware-tuning surface. If open source is a requirement, Jan is the pick of the desktop class.
- Does Jan have an API other apps can use?
- Yes — it runs an OpenAI-compatible server on localhost:1337, so BYO-model tools can target Jan exactly as they would Ollama. Combined with stable MCP support (since v0.6.9), Jan works as a local backend for agentic setups, not just a chat window.
- Can Jan use cloud models too?
- Optionally — bring your own keys for OpenAI, Anthropic, Mistral, Groq and others (plus Anthropic-compatible custom providers), alongside local models in one interface. Local stays the default and the point; cloud is an add-on, and there's no hosted Jan service.
Related
- LM StudioA desktop app for discovering, downloading, and running open-weight LLMs locally with a GUI and a local OpenAI-compatible server.
- OllamaAn open-source tool to run open-weight LLMs locally with a single command, including a local OpenAI-compatible API.
- Llama CppThe C/C++ inference engine that made local LLMs possible — GGUF quantization, every GPU backend, and an OpenAI-compatible server, with no dependencies.
- Best Tools for Running LLMs Locally in 2026The local LLM stack, ranked by job: Ollama for serving tools, LM Studio and Jan for desktop exploration, llama.cpp for control, vLLM when it's real serving.
- QuantizationQuantization shrinks a model by storing weights in lower precision (8-, 4-, even 2-bit) — cutting memory and speeding inference at a small accuracy cost.
- MCP (Model Context Protocol)MCP is the open standard for connecting AI models to external tools and data: write one server, and any MCP client — Claude Code, IDEs, agents — can use it.