# MyAi release notes ## MyAi 0.2.8 — Current Windows download Integrates the website installer-update path with the newer desktop interface and AI-selection work from main. The website advertises the versioned 0.2.8 installer; the original 0.2.7 release remains unchanged. The update-button and saved-state checks passed on the merged app. This remains an unsigned Windows development preview with the installation limits described below. ## MyAi 0.2.7 — Website installer updates Settings → App updates → Download latest version opens the official website. Download the Windows installer, close MyAi, and keep the existing installation type and folder. No uninstall is required. App identity and saved-data locations remain unchanged. The website includes update instructions, signing status, and the installer checksum. The Windows x64 installer is unsigned. Windows security policies may block it. Real Electron tests verified the website button and synthetic saved-state persistence across a restart; installed-version replacement remains unverified on this computer. This release does not change Android or iPhone builds. ## 0.2.6 — Desktop updater implementation; activation pending Added in-app update status, background checks/downloads for configured signed Windows releases, differential-download support, and explicit restart installation. Drafts and active work prevent unsafe restarts. Added a signing-required release configuration and owner release instructions. The live update feed, signing credentials and a real signed-version-to-signed-version upgrade remain external activation requirements; development previews show this honestly and do not fetch updates. No mobile OTA updater or new public binary release was enabled. ## 0.2.5 — Web when helpful Desktop and native phone chat can retrieve Bing search-result summaries for current questions and explicit web requests. Answers remain on the device (or the explicitly paired desktop). Replies include clickable source links; offline settings, cancellation, relevance filtering and unavailable-search messages are implemented. Search sends only the current query, not saved history, attachments or personal notes. Search results are untrusted summaries, not full articles or a guarantee of freshness. Very long prompts, attachments, code blocks and obvious credentials are not searched. No paid provider, API key or cloud inference was added. macOS/iPhone execution still requires Apple validation. ## MyAi 0.2.4 — simpler everyday use Automatic setup chooses a suitable free local AI, reuses existing downloads, and loads it without a model decision. The ordinary chat interface hides token counts. Desktop and mobile helpers start with a name and a description; extra access and technical controls are optional. Simple desktop helpers open as normal conversations. Website navigation, getting-started help, and local preview downloads follow the same flow. The local Windows installer and Android APK are in `artifacts/simple-experience/`. These remain unsigned/development-signed previews. Mac, iPhone, physical-phone validation and public release approval remain separate. Previous release notes below are historical. ## Historical HOLD 0.2.0 Local validation build — 2026-10-15. Built by Elpmis Software. No public release has been published. ## Included - Real local text and photo inference, CPU/Vulkan/NVIDIA CUDA options, compatible model recommendations, verified/resumable downloads, local GGUF import and separately approved model revision checks. - A refined desktop workspace with per-chat roles/instructions/response settings, editable messages, branches, side-by-side comparison, text/photo drop and attachments, long-history paging, rich Markdown/code copying, exports and built-in help. - Local Windows dictation, WAV transcription and spoken replies using installed speech resources. Review transcripts before sending; no cloud fallback. - Explicit personal/project memory, semantic project retrieval and encrypted saved-history search with current-source validation and deletion invalidation. - Agents with scoped project/MCP/web/memory/test permissions, bounded delegation, durable task queues and visible activity; exact review for edits and external calls. - Durable multi-file checkpoints and recovery, reviewed Git operations, and isolated local Node tests with resource limits, no network, read-only copied sources and verified cleanup. - Private desktop/phone pairing with pinned TLS, mutual request authentication, resource budgets and revocation. Separately granted desktop-agent tasks require desktop approval. - Native Android local inference, camera/photo/share/QR integration, offline WAV transcription and spoken responses where supported, conversation controls and corrupt-storage recovery. Native emulator evidence is separate from physical-phone and Apple-platform testing. - Optional explicitly approved remote model connections, a deployable static website, current source documentation, verification evidence and version-specific artifact checksums. - Save-failure recovery preserves unsent text and attachments. Dialog errors, focus restoration and keyboard tabs are accessible; Android response details show local timing measurements, and the narrow-screen keyboard keeps Send reachable at large text sizes. ## Artifact trust and platform scope Windows artifacts are unsigned. Android artifacts use a development signing identity. Hashes identify exact bytes; they do not identify a trusted public publisher. Model weights are separate downloads, not included in the source archive. CPU/Vulkan/CUDA results were observed on the recorded Windows machine. Android evidence comes from an emulator. Native macOS, iOS, physical-phone and other-device validation requires the corresponding hardware/toolchains. Paid provider execution remains disabled without enforceable provider caps and owner authorization. Public marketplace/settlement features remain an architectural roadmap, with no advertised earnings or operational service. No release updater, public hosting, store submission or paid service has been enabled. Small models can fail reasoning, coding or tool tasks. Strict benchmark failures, dependency advisories and platform limitations are retained in the evidence. Review [verification](/reference/docs/VERIFICATION.md), [requirements](/reference/docs/REQUIREMENTS.md), [security](/reference/docs/security.md) and [owner setup](/reference/OWNER_SETUP.md) before distribution. The earlier 0.1.0-alpha.1 artifacts remain preserved as a historical baseline; this build has separate evidence and checksums.