Open Minis

New: "Ask Minis" via Siri — say "Hey Siri, Ask Minis" to jump straight into a live conversation. New: Kimi Code sign-in support, plus full OAuth login for MCP servers (with iCloud config sync across devices). New: Codex Fast Mode toggle for OpenAI OAuth models. New: Live Activity (privacy mode) now shows real task progress — position, status, and run time. New: Select any text and tap "Read Aloud" to hear it via Minis TTS. Voice correction learning now uses real conversation context, further improving correction accuracy. New: Max output length raised from 64K to 128K tokens. In-browser downloads got a full UX overhaul with floating progress cards and a downloads panel. Stability: Fixed several stability issues in message streaming, Live Activity restore on iPad, HealthKit data reads, and config loading after a device restart. Reduced stutter during long "thinking" streams and scrolling. Bug fixes: - Markdown tables keep their scroll position while streaming - Fixed duplicate thinking bubbles on retry - Fixed invisible scroll buttons in light mode - Fixed navbar title staying blurred after Face ID unlock - Session list timestamps are properly localized - Fixed in-app browser fetch failures - Fixed Mistral chat compatibility issue - Backfilled translations for ~200 previously English-only strings - Added Claude Opus 5 to the built-in model list - Fixed occasional config or chat history issues after a device restart - Fixed occasional unresponsiveness when checking Now Playing info or using speech synthesis - Fixed regenerating a session title not following the latest model order - Fixed some health metric types occasionally being skipped when reading HealthKit data - Fixed onboarding setup steps missing translations in some languages - Reduced app install size

Voice input & output On-device VAD recording with pluggable ASR/TTS providers (Azure, Doubao, Xiaomi MiMo, system online/offline), fail-over across providers, a redesigned voice panel, and an optional AI correction pass using recent conversation context. App Lock with Face ID Lock the entire app with Face ID, Touch ID, or passcode, with a configurable grace period and a privacy screen in the task switcher. model-use passthrough mode Command-line tool calls can now pass raw request bodies, headers, and custom endpoints straight through to a model provider, with warnings surfaced when fields are silently dropped or downgraded. Thinking level in the nav bar A badge to view and change a session's reasoning effort, with Max/Ultra tiers, a picker, and a slash-menu shortcut. Streaming rendering Messages glide into place with word-by-word fade-in; code blocks animate open as they stream. Model Quick Test Test a model's top-3 modalities concurrently, with results shown side by side. Live Activity Play/pause and mute/unmute narration from the Dynamic Island and Lock Screen; fixes stuck timers and stale state after a session finishes. Compact Messages A "Compact Messages" menu entry and a one-tap auto-compact option on the capacity warning. Also in this update - Long-press the new chat button to start a session with a specific model group - Azure OpenAI mode for OpenAI/Responses providers - Modality icons and priority tags on Model Groups - Editable Context Window for any model - Scroll-to-top button in long chats - Print support for web pages, images, Markdown, and text previews - Copy table as image - Session titles summarize both the first and last turns - Cache hit rate shown in the Token Usage sheet - Claude Sonnet 5 and GPT-5.6 (sol/terra/luna) added; Grok list synced - Multiple fixes to the built-in terminal, improving stability Bug Fixes - Fixes the "..." menu refreshing or resetting mid-streaming - Fixes context-window usage showing stale numbers after editing a model or model group - Fixes English text wrapping mid-word instead of at word boundaries - Fixes a bottom gap in the voice panel after edit-and-send - Fixes a 3-second UI freeze when switching from voice to text input - Fixes voice-template model sync deleting API-fetched audio models on launch - Fixes three app crashes caused by dangling view references - Fixes Markdown code block spacing inconsistencies - Fixes voice input panel gestures conflicting with transcript scrolling and editing - Fixes the chat input bar retaining stale height across re-entry - Fixes terminal crashing after a rootfs reset - Fixes silent playback in the System TTS/ASR path - Fixes thinking not disabled for adaptive-reasoning models set to "off" - Fixes Gemini models failing when a response contains an empty text segment during tool calls - Fixes reasoning mode not working for some third-party models - Fixes retried messages occasionally disappearing from chat history - Fixes Token Usage counting inaccuracies on reloaded sessions - Fixes a crash on iOS 26.5 caused by a removed WebView API - Fixes table columns collapsing or clipping emoji during streaming - Fixes Bluetooth keyboard send leaving the software keyboard on screen - Fixes a memory leak in iCloud sync's per-session cache - Fixes idle CPU sitting at 100% in the session list - Fixes a slow cold start when opening a session - Fixes main-thread hangs during long-message streaming

What's New 1. MCP Integrations — Connect Any Tool to Your AI Minis now supports the Model Context Protocol (MCP), letting you connect external tools and services directly into the agent loop. Add MCP servers via Settings → MCP Integrations, import configs from JSON, or manage them from the shell with the new  minis-mcp-cli  command. 2. SOUL Personality Customize your AI's personality — name, style, language, and character — directly in Settings. Everything is saved to SOUL.md and persists across sessions. The sidebar title, default session name, and typing indicator all reflect your Soul's name. 3. iCloud Sync V2 Upgraded to record-level sync. Providers, Model Groups, and environment variables now sync independently per entry. Fixes issues where deleted configs were resurrected by iCloud, local API keys were overwritten on merge, and active sessions had their content overwritten mid-run. 4. Providers OpenAI: Generate images directly in chat using gpt-image-2 via OpenAI OAuth / Codex instances. New xAI provider with one-tap SuperGrok / X Premium+ login. 5. Memory Toggle A new global memory toggle in Settings controls whether new sessions enable memory by default. Turning it off fully suppresses memory injection and hides memory tools. 6. Tool Calls The agent loop now dispatches multiple tool calls in parallel, making multi-step tasks significantly faster. Long-press any tool capsule to re-run from that exact point, or copy the tool call details. Bug Fixes • Fixed multiple chat crashes including session switch UAF, notification deeplink wake, and iPad sidebar collapse jank. • Fixed generated content being lost after Stop or Retry. • Fixed vision models not seeing images returned by read_image or browser_use. • Fixed concurrent browser_use calls trampling the same WebView across tabs. • Fixed Codex OAuth non-streaming calls triggering HTTP 400 Stream must be set to true. • Fixed orphaned tool_use and tool_result blocks being sent in Anthropic requests. • Fixed reasoning models returning empty text when generating session titles. • Fixed slash menu disappearing in over-input mode and installed Skills not appearing in the list. • Fixed token counts being inflated by providers that report cumulative usage on every SSE chunk. • Fixed compact running in an infinite loop. • Fixed image generation requests timing out at around 120 seconds. • Fixed camera captures being stored as oversized lossless PNGs.

New Features 1. minis-config CLI: agent edits app settings from terminal with audit and revert. 2. Environment variable privacy mode: env values masked in shell output before reaching model. 3. Full minis:// deeplinks: external links jump into Settings and file browser. 4. New Chat onboarding: workspace folders and file browser entry on start. 5. File mention ranking: @ mention ranks closer to real usage. Bug Fixes 1. Empty responses: GPT-5.5 blank replies now show error with retry. 2. Reasoning fallback: fixed early stop from fabricated placeholders across providers. 3. Long-session stability: 200-turn cap with resume; cross-provider tool 400 fixed. 4. DeepSeek V4 compatibility: fixed misflagging as vision-capable. 5. Image attachments: picker/drop images keep original bytes, no JPEG. 6. Streaming scroll jumps: fixed half-screen jump and broken table links. 7. Image preview transition: full-screen image no longer freezes mid-animation. 8. Misc polish: empty state visuals, browser preview sizing, settings link routing. Performance 1. Tool streaming heat: CPU and device heat drop on large tool inputs. 2. Long-chat smoothness: smoother scroll and streaming; logs shrink dramatically.