2.8.4 (2026-08-19)

• Added a new Security settings section for host-side plugin permissions.
• Added separate read and write restrictions for local filesystem access outside the workdir data directory.
• Added per-system Linux, Windows, and macOS system-command whitelists and blacklists, with the whitelist taking precedence when enabled.
• Added shared filesystem and command security checks across filesystem-capable plugins and host-side command execution.
• Added Permission denied tool responses that point to Settings -> Security when filesystem access is blocked.
• Added a Computer Use tab in Security settings.
• Added Halt on potentially unsafe operation, enabled by default.
• Added explicit user confirmation for provider-flagged potentially unsafe Computer Use operations in OpenAI and Google providers.
• Added per-model custom API endpoint and API key settings.
• Improved the Models Editor advanced settings layout and added missing field descriptions and translations.
• Moved application-managed temporary files for audio input, HTML Canvas, Code Interpreter/IPython, and the Transcript tool to the workdir tmp directory.
• Centered the audio input/output level meter in the bottom input bar.
• Fixed the Painter tool restoring the wrong color icon after application restart.
• Fixed escaped Unicode characters in stored and displayed command/tool JSON, including nested tool responses.
• Updated examples and documentation.

2.8.3 (2026-08-16)

• Fixed duplicated attachments being displayed after tool calls.
• Fixed duplicated images in tool chains.
• Fixed attachment paths in additional context so the full filesystem path is no longer sent.
• Fixed URL extraction from responses.
• Fixed reasoning context being appended to responses for local Ollama models such as DeepSeek.
• Added native attachment file upload when supported by the provider.
• Added real-time reasoning rendering.
• Added collapse/expand controls for URLs and attachments lists.
• Added missing settings descriptions.
• Upgraded the image generation plugin to support the newest models.
• Changed attachment archive rendering to display an archive as a single entry.
• Changed tool-chain rendering to display action icons at the end of the chain.
• Set max_total_tokens to 0 by default.
• Updated the context max limit option description.
• Updated documentation and configuration descriptions.

2.8.2 (2026-08-15)

• Added ARM64 AppImage and Snap builds.
• Added support for preserving full archive paths when attaching ZIP files.
• Removed the Send mode: Off option.
• Improved tool call rendering.
• Improved CSS and UI.
• Moved the Clear input on send option to Settings.

2.8.1 (2026-08-13)

• Added LiteLLM as LLM provider - #181, thanks to @RheagalFire.
• Added Forge as LLM provider - #173, thanks to @Yiiii0.
• Added Eden AI as LLM provider - #194, thanks to @MVS-source.
• Added TwelveLabs plugin (video understanding with Pegasus + Marengo embeddings) - #191, thanks to @mohit-twelvelabs.
• Fix: use unique filename per xAI TTS generation to avoid Windows file-lock PermissionError - #192, thanks to @LittleBallOfPurr.
• Fix: image and video URLs formatting in Windows.
• Added an option to hide models from the model list in the Models Editor.
• UI improvements.

// edit: 2026-08-14, 13:40

• Updated AppImages: added arm64 architecture.

2.8.0 (2026-08-12)

• Upgraded all models to their latest available versions.
• Upgraded dependencies.
• Fixed compatibility with the latest models.
• Fixed the Runner agents workflow.
• Fixed the Ollama provider and custom model handling - #169, thanks to @ba2512005.
• Fixed top_p handling and added temperature guards for Claude 4.x Anthropic models - #184, thanks to @robertnama.
• Fixed message handling for Mistral - #185, thanks to @WildGreenRose.

• xAI SDK upgraded to v1.6.1.
• Added video generation support in xAI provider.
• Added new models: grok-imagine-image and grok-imagine-video.
• UI improvements and smoother navigation between tabs and columns.
• Optimized adding attachments. Added options to attach an attachment only once (if the API stores messages on its side), and an option to force attaching an attachment only once.
• Updated libraries.

• Improved handling focus between columns.
• Fixed models debugger update.

• Fixed an issue where an avatar could be overwritten when creating a new preset.
• Fixed an issue where a new context was not created when opening a new tab in the second column.
• Added prompt history navigation to the input field (Ctrl + Up/Down arrow keys).
• Added initial image centering when loading the Image Viewer.
• Added a Mark/Unmark feature to the Notepad widget.
• Added 18 new languages: Arabic (ar), Bulgarian (bg), Czech (cs), Danish (da), Finnish (fi), Hebrew (he), Hindi (hi), Hungarian (hu), Japanese (ja), Korean (ko), Dutch (nl), Norwegian (no), Portuguese (pt), Romanian (ro), Russian (ru), Slovak (sk), Swedish (sv), Turkish (tr).

• Improved realtime audio mode.
• Added xAI provider and Grok support in realtime audio mode.

• Added the xAI Collections remote tool and integrated collections management into the Remote Vector Stores tool.
• Unified the Remote Vector Stores tool into a single tool for all providers.
• Added xAI Grok audio input and output provider (realtime audio support for Grok coming soon!).
• Enhanced the Image Viewer tool with a toolbox featuring Prev/Next and more options.