Jan

Jan lets you download and run supported language models on your computer, chat with custom assistants and expose a local OpenAI-compatible server to other applications. It suits users who want control over model choice and local inference, plus developers integrating AI into local workflows. It also supports optional cloud providers and MCP servers; a local chat app is not a guarantee that every configured feature stays offline.

Mac requirements and installation

The exact 0.8.4 jan-mac-universal-0.8.4.zip release asset is 102,371,540 bytes, verified by the publisher's GitHub asset metadata. The current official Mac guide says Apple Silicon is supported and Intel Macs are not supported, despite the historical universal archive name. Use macOS 13.6 or newer. Metal acceleration is enabled on M-series Macs. The guide recommends roughly 8 GB for smaller 3B models, 16 GB for 7B and 32 GB for 13B, with quantization, context length and other apps affecting fit. Keep at least 10 GB free storage and more for models/caches. These are model-fit guidelines, not promises that every model of that parameter count will load.

Extract the archive and move Jan to Applications; the publisher also offers DMG installation. Download a compatible model separately through the Hub or configure an existing model/provider. The app archive does not include every model. Node, Yarn, Rust and Xcode/MetalToolchain prerequisites in the README are for source builds, not ordinary installed-app usage. Backend/model acceleration availability depends on the selected model and installed backend; CPU-only inference can be much slower.

Features and dependencies

Use local models such as supported Llama/Gemma/Qwen/GPT-oss variants, manage conversations and assistant parameters, or connect a provider with your own credentials. Attachments and document retrieval depend on parser/embedding/model capabilities. Vision and tool calling require models that support them. The default local API address is localhost:1337; keep it loopback unless you deliberately configure and secure remote access. Port conflicts can prevent startup.

License, accounts and cost

Jan is Apache-2.0 licensed by Menlo Research. Local-model use has no Jan subscription or cloud-model usage fee, but downloaded models have their own licenses and use restrictions. Model downloads require network/storage; local inference can run offline after the necessary files are available. Cloud providers, search tools and external MCP services may require accounts, keys and charges. Inspect their pricing and data handling before enabling them.

Permissions, storage and privacy

The official privacy documentation describes opt-in usage analytics, a random identifier and PostHog EU, with preferences selectable at first launch and in Settings. It says analytics do not collect chat prompts or conversations. Cloud models receive the messages you send to them; MCP servers can access or act on data according to their tools and permissions. Trust only intended endpoints, limit filesystem tools and review approval prompts. Model-generated commands or documents need human review.

Version 0.8.4 migrates provider settings from webview localStorage into a backend-managed store and secrets into the OS keyring. Old localStorage remains as a downgrade snapshot; changing keys/settings in 0.8.4 does not update that old snapshot. Rotate credentials deliberately and account for retained copies when downgrading or cleaning a shared machine. Chats, models and logs are stored locally; local storage can still be read by someone with access to the account/device. Inspect logs before sharing them. The separate MCP routing callback documented for contributors is not registered by shipped Jan, and must not be confused with app analytics.

Reliability and risks

Choose models that fit available memory, reduce context length when needed and back up important conversations/settings. AI answers can be inaccurate; tool calls can change files or contact third parties. Running a local model removes a provider from that inference path, but does not sanitize inputs, guarantee secure outputs or make external tools private. Updates may change backend support or configuration. Verify the actual supported Mac guide rather than relying on a downloaded file's name to infer CPU support.

Sources: Tagged README, Official Mac guide, Privacy, Routing callback scope, Exact releases, License.

Newer catalog release

The catalog now includes 0.8.5, 0.8.4 and 0.8.3. The current package download remains the 0.8.4 universal ZIP above; its 102,371,540-byte size must not be interpreted as the size of 0.8.5. Version 0.8.5 bundles an Apple Silicon-only local engine, separates the CLI, enforces Trusted Hosts on publicly bound API listeners and retains conversation prompt caches on disk by default. See the exact release notes for migration details and limitations.