Jan

Overview

Jan is an open-source desktop assistant, licensed under Apache 2.0, that downloads a foundation model on first launch and answers from it locally. Local inference runs through llama.cpp or MLX, and the built-in Hub lists each model with whether it fits the hardware in front of you, including Menlo Research's own Jan-v3-4B, the vision model Jan-v2-VL, Jan-Code-4B and the Jan Nano models at 32k and 128k context. Cloud models sit beside them rather than replacing them: Anthropic, OpenAI, Azure OpenAI, Gemini, Groq, Mistral AI, OpenRouter and Hugging Face are configured with your own key, and any other OpenAI- or Anthropic-compatible endpoint can be added by hand. Around the chat there are Projects that index uploaded files for retrieval, Assistants, MCP servers, and web search and page fetching that work without a key. A local server exposes the same models on an OpenAI-compatible endpoint at port 1337. Two companion pieces are distributed separately: Jan Agent, a command-line coding agent in preview, and Tokamak, a self-hosted router with governance and audit that agents connect to for model switching. The software is free; only the cloud providers you connect cost anything, and they bill you directly.

Best for developers and privacy-minded teams who want a desktop assistant whose default model runs locally, with cloud providers optional.

Jan is the rare assistant that treats a local model as the normal case rather than a privacy setting, and the small things follow from it: search that needs no key, files chunked on your own disk, a server you can point other software at. For anyone whose objection to a hosted assistant is where the documents go, that removes most of the argument without ruling out a frontier model later. The limits are worth knowing before you commit. It lives on a desktop, so there is no phone and no browser tab, and answer quality tracks the machine: a laptop that cannot hold a larger model will feel slower and blunter than a cloud chat, and the fix is either better hardware or a provider key. The parts that would make it a team tool, Tokamak's governance and the agent, are newer than the chat and still carry preview labels, so treat those as things to trial rather than to deploy.

Key Facts

Features

Ecosystem

Desktop app for macOS, Windows and Linux, with Jan Agent as a separate command-line agent and Tokamak as a self-hosted router carrying policy controls and audit trails for organisations. MCP servers attach under Settings, and documented integrations cover Claude Code and OpenClaw

Live web search and sources

Yes, built in and on by default: web_search and web_fetch are advertised to any model that supports tool use, with a keyless Exa endpoint as the default provider, so no API key and no MCP server are needed. Only tool-capable models see the tools

Files and long documents

Yes, files attach to a single message, or go into a Project where they are chunked and indexed for retrieval across every conversation in it. PDF, Markdown, DOCX, XLSX, PPTX and code files are supported, and processing happens on the machine

Images, voice and video

Images and audio attach to a chat when the active model can take them: the image option appears only with a vision-capable model such as Jan-v2-VL, and audio accepts WAV and MP3

Coding help

Jan-Code-4B is available in the Hub for local coding, and Jan Agent runs coding tasks in a terminal with subagents, skills and memory, asking for confirmation before each write and shell command. It is a preview shipped on a nightly channel

Model choice

Yes, and both kinds mix in the same app: local models through llama.cpp and MLX from the Hub, or cloud models from Anthropic, OpenAI, Azure OpenAI, Gemini, Groq, Mistral AI, OpenRouter and Hugging Face, plus any OpenAI- or Anthropic-compatible endpoint

API

Yes, a local OpenAI-compatible server at http://127.0.0.1:1337/v1, authenticated with a bearer key, with configurable host, port, path prefix and CORS

Integrations

• OpenAI
• Anthropic
• Google Gemini
• Azure OpenAI
• Groq
• Mistral AI
• OpenRouter
• Hugging Face
• llama.cpp
• MLX
• Model Context Protocol
• Claude Code
• OpenClaw
• Exa

Deployment and Support

Deployment

• On-premise

Support

• Community forum

Onboarding

• Documentation and knowledge base

Company size

• Freelancers
• Small and mid-sized
• Enterprise

Suitability

The table combines the industries the vendor addresses with the company sizes the tool is aimed at. An empty cell means the combination is not stated, not that it is ruled out.

Pros and Cons

Pricing

• Free

  $0

  The whole desktop app under the Apache 2.0 licence. Local models run on your own hardware; cloud models are billed by the provider whose key you add.

Best For

• Technology

Alternatives and Comparisons

FAQ

Does Jan work without an internet connection?

Yes, for local models. That is the default: a model is downloaded on first launch and runs on your machine with no API key. Cloud providers and web search need a connection.

Can Jan use GPT, Claude or Gemini as well?

Yes. Anthropic, OpenAI, Azure OpenAI, Gemini, Groq, Mistral AI, OpenRouter and Hugging Face are set up in Settings with your own API key, and any other OpenAI- or Anthropic-compatible endpoint can be added. Local and cloud models can be mixed in the same app.

Can other programs use the models Jan has downloaded?

Yes. Settings has a local API server that serves an OpenAI-compatible endpoint at http://127.0.0.1:1337/v1, so anything written against the OpenAI API can point at it. Host, port, prefix and bearer key are configurable.

What are Jan Agent and Tokamak?

Jan Agent is a separate command-line agent for coding work, currently a preview on a nightly channel; it ships no inference engine of its own and calls a provider. Tokamak is a self-hosted backend for organisations that routes between local and cloud models and adds policy controls and audit trails.

What users say

Reviews

Share your experience

Write a Review