Jan

Jan is *ChatGPT that runs entirely on your own machine* — an open-source (MIT) desktop app that downloads and runs local LLMs with a clean chat UI, so your prompts and conversations never leave your computer.

Background. Released in 2023 and *MIT-licensed*, Jan wraps the llama.cpp inference engine in a friendly desktop application: model discovery and download, chat history, and an *OpenAI-compatible localhost server* so other tools can talk to your local model as if it were OpenAI's API. It's cross-platform and explicitly aims to be "the open-source Ollama, with a UI." That fully-local, permissively-licensed design is why it earns an A in /ai — it's the most approachable way to get a private, offline assistant running.

What you trust. Yourself and your hardware. Because Jan runs models *entirely on your machine*, your prompts, your documents, and the model's responses *never transit a third party* — there's no cloud endpoint logging your conversations, no provider training on your data, no account tying your usage to an identity. The app is *open source (MIT)*, so what it does — and that it isn't phoning home — is auditable rather than promised. The *OpenAI-compatible local server* means you can even point existing OpenAI-API tools at Jan and keep all the inference local. The trust collapses to "code you can read, running on hardware you control."

Operational specs. A *cross-platform desktop app* (Windows/macOS/Linux) that wraps *llama.cpp*: drop in a *GGUF model* (or pick one from the in-app discovery), and chat. It maintains local chat history and exposes an *OpenAI-compatible API on localhost* so editors, agents, and scripts can use your local model as a drop-in for the OpenAI endpoint. UX is deliberately tuned for *non-engineers* — you don't need to touch a terminal to get a model talking. Free and open-source; the only cost is your own compute.

Philosophy. The mainstream LLM experience routes your most sensitive queries — drafts, ideas, medical and legal questions, code — straight to a vendor that logs and may train on them. Jan's premise is that capable AI should be *yours*: run the weights on your own silicon, keep the data on your disk, and let anyone inspect the code that does it. Making it *easy* (a real app, not a command line) is the other half of the philosophy — privacy that only experts can achieve isn't privacy for most people, so Jan trades some power-user flexibility for an interface a non-engineer can actually use.

Grade rationale. A in /ai. The grade reflects fully-local inference (zero data egress), an MIT-licensed open-source codebase, an OpenAI-compatible local API for integration, genuine cross-platform support, and a UX that brings local LLMs to non-technical users. It's a standout privacy option in the category. The caveats are about the local-model capability ceiling and hardware needs, not the tool's integrity.

Useful when. Use Jan when you want an AI assistant whose conversations *stay on your machine* — confidential work, private questions, offline use, or simply not wanting a vendor to log your prompts — and you'd rather have a polished app than wrangle a CLI. It's also a clean way to give *existing OpenAI-API tools* a local backend: point them at Jan's localhost server and your data stops leaving the box. Ideal for the privacy-minded who aren't ready to run inference from the command line.

Caveats. Local models are bounded by *your hardware* — a consumer GPU/CPU runs smaller models than a frontier cloud service, so set expectations to your machine; the privacy comes with a capability ceiling. Larger, more capable GGUF models need real RAM/VRAM, and running them well takes some tuning. As an actively-evolving open-source project, features and stability move fast (occasional rough edges). And "private" means *you* are now responsible for the model and data on your disk — your device security matters. None of these undercut the A: the ability to run a competent assistant with *zero* data leaving your machine, in an app a non-engineer can use, is exactly what the category needs more of.