• Compressing and image uploads now work for the osx and windows builds
• Various small bugs fixed
• Support for MCP commands was added via std. For example for Intelij MCP Server
• BatchTool was failing to pass data to some other tools
• Tool calling now shows its description text even after it finished running
• Fixed various scenarios where AbortSignal was not working to stop an operation

• Instead of hardcoded 3001 now runs on any available port (will try 3826 as default first as least common, then find other options if still busy)
• Added "Compress" button on top of each chat, that will use Expert model in order to summarize a long conversation history, to save up tokens as it blows up.
• Added support for MCP servers, now you can add many MCP servers, and use them in chats to add more tools
• Additional fixes to the token counter, now it updates realtime even as agentic loop keeps going

What’s Changed

Added

• view tool: Now skips massive binaries, but it will also be able to "look into" images
• macOS packaging: Builds are now universal (x64 + arm64).
• Background typing in the text editor while the model is busy (sending is still disabled until the model is free).

Fixed

• Message Prompt Editor → Block tab: Send button now functions correctly.
• Token counter: Accurate counts; removed erratic jumps, added new claude and gemini pricing.
• Socket uploads: Large PNG attachments no longer crash the socket connection.

• Multiple bug fixes of the day

Added

• Support for Qwen3 using openai base url, and some post processing fro and <tool_call> tags
• Fixed bugs with the auto updater
• Modal dialogs now close with ESC press

Added

• Finally added CI
• Fixes for OSX related issues when running from DMG
• Fixes for Windows related issues

Added

• Chat Context Visibility – system prompts now visible in message context.
• Theme System – complete CSS structure refactor with support for multiple themes.
• Light/Dark Mode – toggle between light and dark themes.
• New Themes – Caramel Latte & Dark Coffee themes added.
• Agent Flavours – create custom agent configurations with system prompt, model and tool overrides per chat.
• Auto-Update Feature – automatic updates for npm installations.

Fixed

• Resolved issue with simultaneous tool calls and content responses.

Changed

• CSS Structure – total refactor for improved maintainability and theme support.

Added

• Prompt Editor – edit prompts as drag-n-drop blocks (foundation for a future system-prompt library).
• New LLM providers: Anthropic (Claude), Google Gemini, Google Vertex AI, and Ollama.
• Custom base-URL support for all providers, bringing the roster to:
  OpenAI, Azure OpenAI, DeepSeek, Groq, Anyscale, Fireworks, Together, Mistral, Perplexity, OpenRouter, Gemini, Vertex AI, Claude, Ollama.
• Settings dialog 2.0
  • Split “Web Fetch” options into its own tab.
  • The models tab now shows provider types.
  • Full CRUD for providers (create, edit, delete).
  • Three independent model-preset slots (aux, main, expert) – each can point to any provider.

Changed

• LLM middleware refactor – nuked the spaghetti; now lean, readable, and provider-agnostic.

Fixed

• First user message is no longer accidentally duplicated in LLM conversations.

Removed

• Dead “shitcode” purged from middleware.

• Fixed critical bugs such as tool calling not working with Access Denied issue
• Changed model prompts to avoid confusing with user/agent roles as well as confusion about replace/edit tool uses
• Removed not workign options from settings dialog temporarily
• Added resizability option to message box
• Fixed system prompt to better follow user instructions
• Agent links now open in new window
• Bashs cripts are now preventable and stoppable