Glance
_
Core trick
Streams macOS screen capture and accessibility-tree changes into the agent's context so it keeps continuous screen state instead of re-screenshotting before each click.
Description
_
README excerpt
# Glance
> Always-on multi-modal perception for AI computer-use agents on macOS.
> **Built solo at Claude Build Day, San Francisco, 2026-06-13.**
Every existing AI computer-use system — OpenAI Codex Computer Use, Anthropic Claude Computer Use, `trycua/cua`, `ghostwright/ghost-os`, `OpenCodexLabs/open-codex-computer-use` — runs the same broken loop for every single action:
```
list_apps → list_windows → walk AX tree → screenshot → guess coords
→ click → screenshot → re-reason → repeat
```
That's **1.5–3 seconds of perception tax** before every click. The LLM has zero continuous awareness of the screen — it re-derives the world from scratch on every action.
Humans don't work like that. We *glance* and see everything at once — every element, every change, continuously, with zero look-up time. **Glance gives AI agents the same thing.**
---
## Architecture — five push-driven perception streams fused into one always-current world model
```
┌─────────────────────────────────────────────────────────────────────────┐
│ GLANCED (Swift daemon — always running) │
│ │
│ Five perception streams run continuously, in parallel, never block: │
│ │
│ ┌──────────────────────────┐ ┌──────────────────────────────┐ │
│ │ 1. AXObserverHub │ │ 2. VisionHub (SCK streams) │