Aya exposes 6 tools and 3 resources over JSON-RPC 2025-03-26. Connects to external MCP servers and merges their tools into every reasoning turn. .mcp.json discovery config ships for Claude Code and Cursor auto-connect.

Short turns (under 10 words, no reasoning signals) route automatically to DiffusionGemma 26B — 4x faster than autoregressive Gemma 4 at 1107 tok/s. Fallback chain: DiffusionGemma → Gemini Flash → Sonnet.

Aya speaks as tokens arrive, not after the full response is done. chunkForSpeech() splits the growing text buffer at sentence boundaries. Text and audio advance in lock-step.

Zero-dependency Web Audio API tones for app open/close, message send/receive, voice start/stop, and error. No audio files. Respects prefers-reduced-motion.

Aya proposes mutations of her own system prompt, A/B tests on the same eval corpus, and promotes only on a meaningful win with a regression guard. Persisted in Supabase. Daily cron via vercel.json.

GraphRAG edge scoring prevents duplicate memory writes (arXiv:2606.01444). Compression ratio surfaced in the Metacognition Coach. Search-class writes routed to the graph.

Eliminated three recurring NotFoundError insertBefore crashes. Root cause: .remove() on React-owned nodes racing the reconciler. Error handler now uses a portal div outside the React root.

Aya leads with warmth before competence (Fiske SCM). Avoids the competent-but-cold envy quadrant. Wired into the chat route via warmthBlock.

Five-dimension evaluation: capability, cost, latency, auditability, workflow lift. Baseline vs Aya Pipeline vs Reasoner. Public receipts at /api/aya/audit.

Mathematical intent classifier + Cognitive Spine role detection selects the right agent and model for every turn. One source of truth in select-model.ts.

Add inputMode field to the chat request body so voice-originated turns route to the speed tier regardless of word count — letting longer voice commands still benefit from the 4x speed advantage.

Stream MCP tool results incrementally rather than waiting for the full response. Reduces perceived latency for filesystem and database tool calls.

Surface skill confidence trajectories in the Developer Hub — EMA over time, decay events, retirement signals.

DiffusionGemma running in the browser via WebGPU for fully offline turns. Replaces the current rule-based offline brain with an actual small model.

Shared workspace state with per-user memory isolation. Real-time presence and collaborative skill crystallization.

One-shot Miso voice cloning via AYA_VOICE_ID. The infrastructure is built; the UX for capturing a voice sample needs design.

Export eval receipts as signed JSON blobs for external audit pipelines. Verifiable by third parties without access to the instance.

Opt-in cross-instance memory federation. Skills and preferences can propagate between Aya deployments under explicit user consent.

iOS and Android clients with push-based proactive speak, background agent sessions, and native audio context management.

Use the skill library as fine-tuning data to produce a small specialized model trained entirely on proven procedures.