What it is
Jan is a desktop app from Menlo Research for downloading and chatting with open models on your own computer, on Windows, macOS and Linux. It runs models through a bundled llama.cpp engine and can also connect to cloud providers such as OpenAI, Anthropic, Mistral and Groq with your own keys.
It serves an OpenAI-compatible API at 127.0.0.1:1337 so other tools can use your local models, and supports MCP tools. Recent versions add a Jan agent with a sandboxed shell tool and built-in subagents for research, review and debugging. Menlo also trains small models for it, such as the 4B-parameter Jan-nano.
The latest release is v0.8.7, from 20 October 2026. Local models do not run on Intel Macs, a limitation carried since v0.8.5.
Our review
Jan is a strong pick for one person who wants local models in a polished app with nothing to sign up for. It has about 45,000 GitHub stars and ships often: v0.8.6 on 9 October 2026 and v0.8.7 eleven days later.
Unlike LM Studio, the whole app is Apache 2.0, and analytics are opt-in: the privacy page says nothing is collected until you agree, and then only active-user and retention counts tied to a random ID, processed by PostHog in the EU. Chats sent to a cloud model go to that provider under its terms.
The product is drifting from chat app towards agent. v0.8.7 added built-in subagents and a switch to turn off the shell tool, which is on by default where the sandbox and model support it. The local API binds to localhost but needs no key unless you set one, has CORS enabled by default, and the docs warn against binding it to 0.0.0.0 without a key. The release also removed a hosted Exa search server that v0.8.3 had added to the default MCP list, a reminder to review what tools a local model is quietly given.
Hardware is the ceiling. The README suggests 8 GB of RAM for 3B models, 16 GB for 7B and 32 GB for 13B on a Mac, and Intel Macs cannot run local models at all. With 614 open issues and pull requests it is busy rather than settled, and it remains a single-user desktop app where Open WebUI serves a whole team from one box.
Where it earns its place
A solicitor handling privileged client files on a 32 GB MacBook
Summarise and query documents without sending them to a cloud provider
Run a 7B to 13B open model locally in Jan with no account and analytics left off. The cost is the laptop; the trade-off is answer quality below frontier cloud models on hard reasoning.
A developer prototyping against the OpenAI API on a train with no signal
Keep coding against an OpenAI-style endpoint offline
Point the app at http://127.0.0.1:1337/v1 and Jan serves a local model through the same API shape. Set an API key before exposing it to the network, since none is required by default.
An IT manager at a 40-person design studio
Offer staff a private chat tool without a per-seat subscription
Jan is free to install from the Microsoft Store, Flathub or direct download, but each install is separate, so there is no central admin, shared history or policy control. For managed rollout a server tool such as Open WebUI fits better.
Before you commit
- Setup
- Download the installer for Windows, macOS or Linux (also on the Microsoft Store and Flathub), then pick a model from Hugging Face inside the app; the first chat takes minutes plus the model download.
- Integrations
- Local models via llama.cpp; cloud models from OpenAI, Anthropic, Mistral, Groq and others with your keys; MCP servers; an OpenAI-compatible API for other apps.
- Data and privacy
- Apache 2.0. Analytics are opt-in, sent to PostHog EU, and exclude chats and model choices; local chats stay on the machine, cloud chats go to the provider. The local API needs no key by default and has CORS on.
- Cost at scale
- Free, with no paid tier. Cost is hardware: Menlo suggests 16 GB of RAM for 7B models and 32 GB for 13B on a Mac, plus your own API spend for any cloud models.
- Limits
- 614 open issues and pull requests; no Intel Mac support for local models; no multi-user features; auto-compaction for remote models needs the context size set by hand.
What's good
- Fully open source, Apache 2.0, no account
- Analytics are opt-in and limited to usage counts
- OpenAI-compatible local API on port 1337
- Mixes local llama.cpp models and cloud APIs
What's not
- Single-user desktop app, no team features
- Local models do not run on Intel Macs
- Agent shell tool is on by default
- Local API has no key unless you set one
Best for
Privacy-minded individuals who want to run open models offline in a desktop app.
Alternatives
Jan: common questions
- Is Jan worth it?
- A fully open, account-free desktop home for local models that has quietly grown into an agent, so its defaults now matter more than its looks. We rate it 4.2 out of 5.
- How much does Jan cost?
- Free and open source (Apache 2.0). No paid tier; cloud models use your own API keys.
- What are the downsides of Jan?
- Single-user desktop app, no team features; local models do not run on Intel Macs; agent shell tool is on by default; local API has no key unless you set one.
- Who is Jan best for?
- Privacy-minded individuals who want to run open models offline in a desktop app.
- How much does Jan cost for a team or in production?
- Free, with no paid tier. Cost is hardware: Menlo suggests 16 GB of RAM for 7B models and 32 GB for 13B on a Mac, plus your own API spend for any cloud models.
- Is Jan safe for company data?
- Apache 2.0. Analytics are opt-in, sent to PostHog EU, and exclude chats and model choices; local chats stay on the machine, cloud chats go to the provider. The local API needs no key by default and has CORS on.
- What are the best Jan alternatives?
- The closest alternatives we have reviewed are AnythingLLM, Cherry Studio and Open WebUI.
Answers checked Oct 23, 2026.
Make Jan? Show this rating on your site
