Migration

Settings and credentials now live in a backend store, not browser localStorage.

Provider settings and credentials previously persisted in the webview's localStorage. In 0.8.4, Jan reads and writes them from a backend-managed store instead (secrets go to the OS keyring). Your existing settings and API keys are copied over automatically on first launch - no manual action is needed, and from this version on all new changes are saved to the backend store.

Your old localStorage data is not deleted - it's kept as a snapshot so you can downgrade to a pre-0.8.4 build and still have your previous settings. Note that any changes you make in 0.8.4+ live only in the backend store, so a downgrade sees the older snapshot, not your latest settings. The migration runs once and is guarded by a backend flag, so re-upgrading never re-imports the stale localStorage copy over newer backend data.

Highlights

Native Web Search & Web Fetch

Jan now ships built-in web_search and web_fetch tools backed by a native tauri-plugin-websearch, so models can search the web and fetch page content without configuring an external MCP server. The tools are available to the agent loop like any other built-in.

Per-Model Chat Template kwargs for llama.cpp

You can now set chat-template keyword arguments per model in llama.cpp router mode, so template variables (for example reasoning or tool-format switches a model's template exposes) can be pinned per model rather than applied globally.

Token Counter for Remote & MLX Providers

The token counter now works for remote and MLX providers, not just local llama.cpp - so you can see token usage during chat across provider types. This landed alongside a batch of related chat and inference fixes.

Interim Text Folding into the Reasoning Trace

A new toggle lets you fold interim narration text into the reasoning/chain-of-thought trace instead of showing it as loose prose between steps, keeping the transcript compact.

Bug Fixes

Chat & Streaming

• Stopping a generation mid model-load now preserves AbortError identity, so the stop is handled cleanly instead of surfacing as an unexpected error.
• Custom SSE event types in OpenAI-compatible streams are now ignored rather than breaking parsing.
• Detached assistant replies in branched threads are no longer persisted, so branch history stays correct.
• Numeric assistant parameters supplied as strings are now coerced to numbers.

MCP & Local API Server

• Guarded against a duplicate serve() on streamable-HTTP MCP servers.
• The MCP tool-approval popup no longer hangs; transient approval state was moved off disk.
• Providers now send the anthropic-version header when fetching the model list.

RAG & Rendering

• Embeddings were repaired, redundant chat-model reloads were stopped, and more PDF formats now parse for document indexing.
• Code blocks are now forced left-to-right (LTR) even inside a right-to-left language context.

Desktop & Platform

• macOS: hide-on-close now hides to the tray when the Local API Server is running, instead of quitting.
• Fixed a snake_case mismatch for the ADL enum in get_gpu_usage.

Security

• Bumped handlebars to 4.7.9 in docs to clear an RCE advisory.

Localization

Japanese (ja) translations were updated across common, settings, and terminology.

Engineering

• App plugin versions bumped to 0.8.4.
• Centralized the llama.cpp extension lookup with a scoped-name fallback.

Update Jan or download the latest (opens in a new tab).

For the complete list of changes, see the GitHub release notes (opens in a new tab).