On‑Device AI:Private Workspace

Version 1.48.0 adds AI Browser, chats that survive interruptions, MCP servers, and much more control over your models and memory. • AI Browser: Browse with tabs, profiles, history, downloads, and restored sessions, then ask your current model to read the page, help you with it, or act on it. The app tells you what a site can see and asks before anything changes outside the app. • Browser tasks that finish: Switch models before a task, follow a clearer timeline of what the agent did, and resize the task pane on Mac. Page reading and action checks improved, the app notices when a task is going in circles, and results are better on long or complex sites. • Chats that survive interruptions: A reply keeps going and reconnects if you navigate away. Queue your next prompt with Send Next or Stop & Send, and compare several models against the same frozen input. • Agent Profiles, schedules, and MCP: Create agent workspaces that stick around, run them when you want, or put supported work on a schedule. Connect remote MCP tool servers and install reviewed AI Skills that cannot run code, while you stay in charge of which tools are on. • Settings for each agent: Give individual Flow Agent steps and Subagents their own temperature and context window, or let them inherit your conversation defaults. • One place for model storage: See how much space chat, embedding, speech, and voice models use. Recover interrupted downloads and safely review duplicate, legacy, incomplete, or orphaned files. • Global Memory you can manage: Search, filter, pin, edit, delete, and undo memories in their own settings area. When the AI wants to change a memory, it now asks the same way tools do. • Faster recovery: Reload your last chat model in one step, and keep a conversation ready when reading aloud borrows a speech engine or restores a model afterwards. • Help, settings, and tools: App Help understands plain questions and comparisons better, support drafts open in more email apps, Settings sections respond faster, and local GGUF and MLX models recognize tool calls more reliably.