OBJECTIVE-FIRST AI OPERATING SYSTEM
Restoring workspace
I°
Loading...
This is a working browser build, not a finished operating system. Core surfaces can be used and evaluated now; others remain experimental, provider-dependent, or under active hardening. This page distinguishes those states.
50+ apps, infinite canvas, widget system, pan/zoom, drag-and-drop. Stable.
Web Speech API recognition, sentence-by-sentence streaming TTS, synchronous mute control, iOS audio unlock. Sound effects via Web Audio API with prefers-reduced-motion support.
Supabase-backed with RLS on all tables. Discovery gate prevents duplicate writes. Local-first cache with cloud sync on reconnect.
JSON-RPC 2025-03-26 over HTTP/SSE. 6 tools, 3 resources. Production endpoint at /api/mcp. .mcp.json discovery config for Claude Code and Cursor.
Strategy Auction routing across conversation (Sonnet), reasoning (Opus), vision (Gemini Flash), speed (DiffusionGemma), micro (GPT-mini). Env-overridable. KNOWN_GOOD_MODELS guard prevents injection.
Five-dimension CAISI-inspired eval. Public receipts at /api/aya/audit. Autoresearch loop promotes champion system prompts on a meaningful win.
The insertBefore NotFoundError crash was fixed. Error handler uses a portal div outside React's root. Initial loader hides via CSS only.
Aya connects to external MCP servers and merges their tools into reasoning turns. The implementation is correct but real-world testing with diverse MCP servers is ongoing. Set AYA_MCP_SERVERS to activate.
Routes short turns to DiffusionGemma (1107 tok/s). The model id google/diffusion-gemma-26b may not yet be available on all Vercel AI Gateway regions — fallback chain to Gemini Flash is automatic.
Daily 04:00 UTC cron promotes champion system prompts automatically. Requires CRON_SECRET, AUTORESEARCH_OWNER_USER_ID, and SUPABASE_SERVICE_ROLE_KEY. Manual promotion via /research always works.
Miso edge/cloud TTS is the primary speech engine when NEXT_PUBLIC_MISO_EDGE_ENDPOINT or NEXT_PUBLIC_MISO_CLOUD_ENDPOINT is set. Falls back to Web Speech API automatically. One-shot voice cloning via AYA_VOICE_ID is built but the UX for capturing a sample is not yet exposed.
Skills crystallize and back-propagate correctly. The confidence trajectory visualization in Developer Hub is in-progress.
There are pre-existing TypeScript strict-mode errors that do not block the build (tsc is non-blocking in the assess pipeline). They are tracked and being resolved incrementally.
Three voice migration tests assert v2 keys in an environment where the migration flag is null. Not a production issue — the migration runs correctly in the browser. Fix is tracked.
The Next.js DevTools sometimes shows a yellow 'stale' Turbopack indicator after HMR. This is a cosmetic Turbopack dev-mode issue, not a production concern.
The best way to give feedback is to open an issue on GitHub or use the feedback surface inside the OS (say "Send feedback" to Aya). For security issues, email directly — do not open a public issue.