Replies now stream in as they're written, plus a handful of fixes. • Streaming replies: answers appear as your model writes them instead of all at once. • Add a folder from the chat: in Agent mode, the paperclip menu now has "Add Folder for Agent" so you choose where it works. It can also create new folders, with your OK first. • Web search follow-ups fixed: asking again about the weather or the news now runs the search instead of showing raw text. • Cleaner replies: hidden reasoning text from some Gemma models no longer shows up. • Drag and drop images from Finder onto the message box to attach them (Mac and iPad).

Web search that just works, sources you can check, and a privacy check before anything leaves your device. • Web search with no setup — Agent mode now searches the web out of the box using DuckDuckGo. No API key needed. Add a Brave Search key if you want a more reliable provider. • Sources under answers — when a reply used web results, the sources it actually drew on appear beneath it as tappable links. • Privacy check — if a web search or an MCP tool call contains an email address, phone number, card number, or API key, CereVault tells you what it found and where it's going, and asks first. • Smarter chat titles — your own model now names each conversation ("Pasta Dough Troubleshooting") instead of using the first few words you typed. • Keys stay in the Keychain — your Brave key and MCP server tokens are now stored in the system Keychain. Existing ones move over automatically. • Fix: email addresses and links in your own message bubbles are readable again.

Agent mode arrives, plus workspaces, connections, and automations. • Agent mode — an opt-in agentic layer over plain chat. Your local model can call tools in a request → act → respond loop: web search (your Brave key), on-device memory, and file access for folders you grant. Writing files is off by default and every change asks first. Tool activity shows inline, and a header info button reveals exactly what the agent can reach. • Multiple workspaces — create, rename, and switch between workspaces, each with its own tools, memory, folders, permissions, model, and chats. Your existing setup migrates into a Default workspace automatically. • MCP servers — connect external Model Context Protocol tool servers to a workspace; their tools join the agent. Every third-party tool call is confirmed first. HTTP/SSE transport. • Automations — schedule a saved prompt to run every few minutes, hourly, or daily; CereVault runs the agent and saves the result as a chat. Read-only by default. • CereVault Serve support — a third local server type alongside Ollama and LM Studio, auto-discovered on your machine. • Refinements — auto-refreshing model list, drag-and-drop image attach, tap-to-zoom thumbnails, adjustable text size, native text selection, and Copy / Copy as Markdown on any message.

What's New Voice dictation. Tap the new microphone in the message field to dictate your message. Uses Apple's on-device speech recognition where supported, so audio stays on your machine. Tap again to stop, or pause for a beat and it stops itself. The live transcript appends to whatever you've already typed, so existing text isn't clobbered. Drag and drop images. Drop image files from Finder, Photos, Preview, or any browser straight onto the chat input to attach them. The field outline highlights when an image is over it. Vision-capable models can analyze the image right away. Tap to view images full screen. Tap any attached image in a chat to open it big with a dimmed backdrop and a tap-anywhere-to-dismiss overlay. CereVault Serve support. CereVault now auto-discovers CereVault Serve instances on your local network alongside Ollama and LM Studio. Prefers them as the default when present and surfaces them at the top of the model picker. Cleaner copy actions. Each chat message now has small Copy and Copy as Markdown icons inline with the sender label, replacing the old long-press menu. After a copy, the icons briefly swap to a green-check confirmation, then revert. Long-press and click-drag now hand off to native text selection again, so you can pull a word, sentence, URL, or code snippet directly out of any reply. Adjustable text size. Cmd+= grows the app's text, Cmd+- shrinks, Cmd+0 resets. The choice persists across launches. Smarter model awareness. The model list refreshes automatically after each chat reply and on chat reopen. If you swap models on the server side, the chat header catches up without needing a manual refresh. Other improvements. Lighter, classier ellipsis icon on chat list rows Demo Mode toggle removed from the sidebar (still available via a five-tap gesture on the CereVault logo, documented in App Review Notes) Chemistry-style equations wrapped in dollar signs now unwrap correctly Common LaTeX commands render as Unicode symbols instead of raw backslash text Various small UI polish and bug fixes