These are client apps only. You need a running Glass Keep server to use them.
On-demand local AI
- Replaces the old Llama 3.2 / Transformers.js / ONNX generator with Qwen3.5 0.8B Q4_K_M and a pinned llama.cpp CPU worker.
- Downloads a verified 533 MB model only on the first question, then reuses the cached file in your persistent data volume. Questions and note excerpts stay on your server.
- Starts the AI worker only for questions and stops it after 60 idle seconds. The process exits; the model stays on disk. Active inference still uses CPU and RAM, and cold loading is not instantaneous.
- Uses a Unicode-aware SQLite FTS5 index and query expansion from the same model to retrieve authorized own/shared notes, including archived notes. No second embedding model, cloud AI key, vector service or extra container is required.
- Shows progressive answers, downloading/loading/searching/writing states, and numbered source-note links. Close the panel to cancel. Small-model answers can still be wrong: verify the linked notes.
- Bounds context to 4,096 tokens and runs one question at a time to limit resource use. Private temporary results expire; permission checks prevent reading another account's job or revoked sources.
- English, Hindi and French controls remain available, with compatibility for older Mac WebViews.
Server/web 1.0.12 · Native clients 1.0.8. Update both for progressive answers and source links. Existing clients can still receive complete JSON answers.
Update
Docker: nikunjsingh/glass-keep:1.0.12 or nikunjsingh/glass-keep:latest (AMD64/ARM64). Pull and recreate in Portainer, keeping your data volume and current settings, including TRUST_PROXY=1 for Nginx Proxy Manager. No new mandatory environment variable is needed.
Enable Settings → Local AI Assistant, enter a question, and press the sparkle button or Enter. The first model download needs internet access and free disk space. AI runs on the server, not the client devices.
The obsolete data/ai-cache is no longer used but is preserved during upgrades. See the local AI guide for safe cache cleanup, resource limitations and optional timeout/CPU settings.
Downloads
Attached: Windows x64/ARM64 installers and portable executables, signed Android ARM64 APK, Apple Silicon/Intel macOS DMGs, and SHA-256 checksums. Android retains the existing signing identity for in-place updates. Windows installers do not have a commercial publisher certificate; portable clients require WebView2. Mac clients require macOS 13+, use an ad-hoc signature and are not Apple-notarized; first launch may require Privacy & Security → Open Anyway.
The native clients still require your server. Cached notes and saved offline changes work as before; server AI needs a reachable server.
Verification
API tests cover indexing existing notes, Unicode, edit/delete tracking, isolation, ignored forged client notes, partial answers, cancellation, failure recovery and legacy clients. Browser tests exercise progressive answers, source navigation and cancellation through both web fetch and native IPC, including missing newer AbortSignal APIs. Existing editor, sharing, offline, export, translation and startup regressions pass. Actual-model Docker checks verify answers, cached reload and idle process shutdown.