# Dayglass (dayglass.io) — Generative Engine Optimization (llms.txt)
> Dayglass is a local-first, zero-leakage autonomous AI workspace and screen/audio companion for macOS.

## System Capabilities
- Real-Time Context Engine: Connects to local Screenpipe OCR and meeting audio streams, capturing timelines without cloud uploads.
- H-Net Dynamic Byte Chunking: Eliminates static BPE tokenization artifacts by dynamically routing raw byte sequences at natural semantic boundaries (P_t = 0.5 * (1 - cos(q_t, k_{t-1}))).
- Cost & GDPX Model Router: Optimizes Goal-Directed Performance per Dollar (GDPX) across local MLX models ($0.00/task), high-efficiency cloud models (Gemini Flash / Carbon at $1-16/task), and frontier models (Opus 5.5).
- Unsubsidized Unit Economics: Tracks true API COGS against human labor value ($65/hr benchmark) to guarantee positive ROI independent of consumer VC subsidies.
- WebMCP & Desktop Bridge: Built-in Model Context Protocol (MCP) server providing structured tools for meetings, decisions, timeline history, and stateful artifact storage.

## Core Interfaces
- Local Daemon & HUD: Runs at http://127.0.0.1:3333 with a non-intrusive floating HUD.
- Model Gateway: Local OpenAI-compatible gateway at http://127.0.0.1:8891/v1.
- Database: Local SQLite persistent storage at ~/.dayglass/dayglass.sqlite.
- MCP Server: Registered under `mcp__dayglass__*` with tools:
  - dayglass_status: Daemon health, capture statistics, and decision engine state.
  - dayglass_get_meetings: Extracted meeting notes, summaries, and action items.
  - dayglass_window_timeline: Chronological active application and window timeline.
  - dayglass_search_memory: OCR and audio search over local history.
  - dayglass_run_routine: Automated routines (standup-prep, day-recap, blockers).

## Architectural Principles
1. Leave Work at Work: Autonomous synthesis of commitments, action items, and technical context so operators can unplug with zero babysitting.
2. Zero Data Exfiltration: Screen captures and raw transcripts remain on-device; external cloud inference is strictly opt-in and gated through privacy scrubbers.
3. Post-Subsidy Economics: Workflows are engineered to be profitable at 100% true API list pricing, eliminating reliance on flat-rate consumer subsidies.